2012 ACM Subject Classification Theory of computation → Proof complexity
Keywords and phrases QBF, CAQE, CDCL, resolution, proof complexity, simulations, lower bounds
Related Version Conference Version : https://doi.org/10.4230/LIPIcs.SAT.2026.7 [17 ]
Funding Olaf Beyersdorff : Carl-Zeiss Foundation and DFG grant BE 4209/3-2
Acknowledgements We thank Leroy Chew for asking a question at SAT’25 on the proof-theoretic strength of CAQE, which led to this work. We used a large language model (Claude) as a writing assistant for this article; all results and proofs were checked by the authors.
Introduction
Quantified Boolean Formulas (QBF) extend the satisfiability problem (SAT) by quantification over Boolean variables. This increases the complexity of the problem (from NP to PSPACE), the versatility of expressions (many problems are more succinctly and more naturally encoded into QBF), and the range of applications [37 ] .
Different from SAT, where conflict-driven clause learning (CDCL) is the ruling algorithmic paradigm, there are a number of competing approaches to QBF solving. Quantified conflict-driven clause learning (QCDCL) lifts the CDCL paradigm [29 ] to QBF [41 , 14 ] . QCDCL can be further enhanced by dependency schemes [31 ] and dependency learning [30 ] and is implemented in a number of QBF solvers (Qute [32 ] , DepQBF [28 ] ). A second approach comes through expansion of universal variables in the counter-example guided abstraction refinement (CEGAR) framework, implemented in RAREQS [25 , 26 ] . CAQE (Clausal Abstraction for Quantifier Elimination) is a third algorithmic paradigm, introduced in 2015 [35 , 39 ] and implemented in the eponymous solver. CAQE also utilises CEGAR, but in a fundamentally different way compared to expansion.
With these various and seemingly orthogonal approaches, it is stunning that since 2017 versions of CAQE have won each QBF evaluation in the prenex CNF track, sometimes by quite a big margin [33 , 34 ] . Experimental evaluations of different solvers as in QBF evaluations are an impactful driver for algorithmic progress. A systematic and theoretically concise comparison of different algorithmic solving techniques – in SAT, QBF and further fields – comes through proof complexity [20 , 14 , 13 ] . Proof complexity of QBF has been hugely advanced over the past decade (cf. [3 ] for a survey) and many theoretically clean and appealing QBF proof systems have been introduced together with proof methods to analyse them (e.g. [9 , 6 , 5 , 11 , 10 , 2 ] ).
When aiming to compare different solvers, we need proof systems tightly modelling solvers and such proof systems are typically quite different from rule-based calculi such as resolution [27 , 40 , 23 ] , Frege or strong systems for proof logging [9 , 24 , 21 ] . A rigorous study of this kind has been performed on different versions of QCDCL and their relations to QBF resolution systems, unearthing surprising incomparability results – e.g. between standard QCDCL and Q-resolution [7 ] – as well as identifying potential improvements for solving (e.g. out-of-order decisions [16 , 19 ] or cube learning and pure literal elimination [18 ] ).
Such a tight theoretical model and comparison to other QBF solvers and proof systems is largely missing for the award-winning solver CAQE. An investigation into the proof-theoretic strength of CAQE was started in the paper [38 ] , focusing on the comparison between CAQE and other CEGAR approaches such as RAREQ§, and establishing their incomparability. However, the relation to QCDCL – arguably the other main QBF solving approach – has been open for the past decade.
Our results. We provide a theoretical proof-complexity analysis of the core procedure of CAQE and compare it to QCDCL and Q-Resolution, resulting in the situation depicted in Figure 1. We do not consider potential optimisations of CAQE such as failed assumptions – we plan to address those in future work.
(a) As a first step, we tightly model the basic algorithmic paradigm of CAQE (without failed assumptions) as introduced in [35 ] as a proof system CAQEC o r e ^{\mathrm{Core}} Core such that runs of the basic CAQE algorithm can be translated into CAQEC o r e ^{\mathrm{Core}} Core proofs and conversely, each tree-like proof gives rise to a run of the basic algorithm (Theorem 11). We note that a proof system ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res was introduced in [38 ] that likewise allows one to translate runs of the basic algorithm of CAQE into ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res , but the converse direction fails (Theorem 16 and Theorem 34).
(b) In our main result we compare the core procedure of CAQE to QCDCL and show the possibly surprising result that QCDCL simulates this basic version of CAQE on Σ 3 b \Sigma^{b}_{3} Σ 3 b QCNFs (Corollary 51). This holds after a simple preprocessing step, which makes the formulas fulfil the X T XT X T -property, and in a model where purely propositional CDCL steps are not counted (see below). Furthermore, we show that QCDCL with cube learning is exponentially stronger on selected formulas (Theorem 40). Our theoretical comparison utilises our new C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proof system and the QCDCL proof systems developed in [18 , 7 ] .
As CAQE uses a massive number of SAT calls, we believe that conceptually the fairest comparison is by discounting SAT computations. This approach aligns with large parts of the QBF literature where propositional hardness is filtered out by allowing NP oracles in proofs [12 ] (whereby lower bounds measure “genuine” QBF hardness and become more challenging). For C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core this is fairly simple: we just count a SAT call as one step, whereas for QCDCL this is somewhat more intricate as QCDCL does not use SAT calls, but has all CDCL functionality “built in”. We therefore define a QCDCL model where purely propositional parts of the traces are not counted (imagine, e.g. that QCDCL stumbles upon a hard propositional formula not containing any universal variables).
Further, we show that our proof system C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core is simulated by Q-Resolution (Corollary 39), again in the oracle model and on bounded quantifier alternation. While ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res is known to be simulated by Q-Resolution [38 ] , this result does not discount SAT calls, which means that it does not imply our stronger simulation. Q-Resolution is a weak system and in particular known to be incomparable to QCDCL [18 , 7 , 15 ] .
Conceptually, the simulation order in Figure 1 means that, from a proof-complexity perspective and in the oracle setting described above, the core procedure of CAQE without failed assumptions is weaker than Q-Resolution on QCNFs of bounded quantifier alternation and weaker than QCDCL on Σ 3 b \Sigma^{b}_{3} Σ 3 b QCNFs. Hence its practical outperformance appears to result from its clever and heavy SAT usage (cf. also the discussion in Section 9) as well as from optimisations such as failed assumptions and the reuse of refinement clauses across different calls (cf. the discussion after Theorem 11) – which we did not consider in this work.
Figure 1. Hasse diagram of the simulation order of QCDCL, Q-Res and CAQEC o r e ^{\mathrm{Core}} Core (stronger systems on top). Solid lines denote a simulation together with an exponential separation. Both simulations hold in the oracle model of this paper: for Q-Res with an NP oracle (QN P ^{\mathrm{NP}} NP -Res) on QCNFs of bounded quantifier alternation, and for QCDCL on Σ 3 b \Sigma^{b}_{3} Σ 3 b QCNFs when purely propositional CDCL steps are not counted (and after a preprocessing step that establishes the XT-property); the separation of QCDCL from CAQEC o r e ^{\mathrm{Core}} Core uses cube learning. The dashed line indicates the known mutual exponential separations of QCDCL and Q-Res in the standard (non-oracle) setting.
Technically, we achieve our results by a characterisation of C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proof size by a measure we call C A Q E CAQE C A QE width, which yields lower as well as upper bounds on C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proof size (Theorem 29, Proposition 32). This characterisation gives rise to exponential lower bounds and the separations as well as to the extraction of short decision lists (Theorem 37) and consequently the simulations.
Relation to the conference version. A preliminary version of this article appeared in the proceedings of SAT 2026 [17 ] . The present article contains complete proofs of all results, most of which were only sketched or omitted in the conference version. In the course of this, we have also corrected some errors and inaccuracies of the conference version.
Organisation. After setting up notation in Section 2 we explain the basic CAQE algorithm in Section 3 and formalise it as a proof system C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core in Section 4. In Section 5 we show that C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core is sound and allows extraction of winning strategies. Section 6 introduces our main technical tool: CAQE width that we use to obtain lower and upper bounds and to prove the completeness of C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core . In Section 7 we show that decision lists for winning strategies can be extracted from C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proofs, yielding the simulation of C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core by Q N P -Res \mathrm{Q}^{\mathrm{NP}}\text{-Res} Q NP -Res . § 8 compares C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core and QCDCL. We conclude with a discussion in § 9 .
Preliminaries
Variables x x x and negated variables x ˉ \bar{x} x ˉ are called literals . We denote the corresponding variable as var ( x ) : = var ( x ˉ ) : = x \operatorname{var}(x) := \operatorname{var}(\bar{x}) := x var ( x ) := var ( x ˉ ) := x . A clause is a disjunction of literals. A unit clause ( ℓ ) (\ell) ( ℓ ) is a clause consisting of only one literal. The empty clause ( ⊥ ) (\bot) ( ⊥ ) has zero literals. A cube is a conjunction of literals, sometimes viewed as a set of literals. We define a unit cube of a literal ℓ \ell ℓ , denoted by [ ℓ ] [\ell] [ ℓ ] , and the empty cube [ ⊤ ] [\top] [ ⊤ ] with “empty literal” ⊤ \top ⊤ . If C C C is a clause or a cube, we define var ( C ) : = { var ( ℓ ) : ℓ ∈ C } \operatorname{var}(C) := \{\operatorname{var}(\ell) : \ell\in C\} var ( C ) := { var ( ℓ ) : ℓ ∈ C } .
An assignment σ \sigma σ of a set of variables V V V is a function σ : V → { 0 , 1 } \sigma: V \to\{0,1\} σ : V → { 0 , 1 } , which can also be interpreted as the set { ℓ : var ( ℓ ) ∈ V ∧ σ ( ℓ ) = 1 } \{\ell: \operatorname{var}(\ell) \in V \land\sigma(\ell) = 1\} { ℓ : var ( ℓ ) ∈ V ∧ σ ( ℓ ) = 1 } . We denote the set of all assignments of V V V as ⟨ V ⟩ \langle V\rangle ⟨ V ⟩ . For a cube D D D we write ¬ D \neg D ¬ D for the clause that consists of the negations of the literals of D D D . A clause C C C is satisfied by σ \sigma σ if C ∩ σ ≠ ∅ C \cap\sigma\ne\varnothing C ∩ σ = ∅ . A cube D D D is satisfied by σ \sigma σ if D ⊆ σ D \subseteq\sigma D ⊆ σ , and it is falsified by σ \sigma σ if ¬ D ∩ σ ≠ ∅ \neg D \cap\sigma\ne\varnothing ¬ D ∩ σ = ∅ . A clause C C C that is not satisfied by σ \sigma σ can be restricted by σ \sigma σ , defined as C ∣ σ : = ⋁ ℓ ∈ C , ℓ ˉ ∉ σ ℓ C|_{\sigma} := \bigvee_{\ell\in C,\bar{\ell} \notin\sigma} \ell C ∣ σ := ⋁ ℓ ∈ C , ℓ ˉ ∈ / σ ℓ . Similarly we can restrict a non-falsified cube D D D as D ∣ σ : = ⋀ ℓ ∈ D ∖ σ ℓ D|_{\sigma} := \bigwedge_{\ell\in D \setminus\sigma} \ell D ∣ σ := ⋀ ℓ ∈ D ∖ σ ℓ . We can restrict a clause C C C to a set of variables V V V by C ∣ V : = ⋁ ℓ ∈ C , var ( ℓ ) ∈ V ℓ C|_{V} := \bigvee_{\ell\in C,\operatorname{var}(\ell) \in V} \ell C ∣ V := ⋁ ℓ ∈ C , var ( ℓ ) ∈ V ℓ . We can restrict a cube D D D to a set of variables V V V by D ∣ V : = ⋀ ℓ ∈ D , var ( ℓ ) ∈ V ℓ D|_{V} := \bigwedge_{\ell\in D,\operatorname{var}(\ell) \in V} \ell D ∣ V := ⋀ ℓ ∈ D , var ( ℓ ) ∈ V ℓ . Note the difference between the restriction C ∣ σ C|_{\sigma} C ∣ σ by an assignment and the restriction C ∣ V C|_{V} C ∣ V to a set of variables – it will always be clear from the context which one is meant.
A C N F \mathrm{CNF} CNF (conjunctive normal form) is a conjunction of clauses and a D N F \mathrm{DNF} DNF (disjunctive normal form) is a disjunction of cubes. A C N F \mathrm{CNF} CNF (D N F \mathrm{DNF} DNF ) is satisfied (falsified ) by σ \sigma σ if all its clauses (cubes) are satisfied (falsified) by σ \sigma σ .
We use the notation σ ⊨ ψ \sigma\models\psi σ ⊨ ψ for a propositional formula ψ \psi ψ and assignment σ \sigma σ if σ \sigma σ satisfies ψ \psi ψ . We write φ ⊨ ψ \varphi\models\psi φ ⊨ ψ for propositional formulas φ \varphi φ and ψ \psi ψ if each satisfying assignment of φ \varphi φ satisfies ψ \psi ψ .
A Q B F \mathrm{QBF} QBF (quantified Boolean formula) Φ = Q ⋅ φ \Phi= Q \cdot\varphi Φ = Q ⋅ φ consists of a propositional formula φ \varphi φ , called the matrix , and a prefix Q Q Q . A prefix Q = Q 1 V 1 … Q d V d Q = Q_1V_1 \dots Q_dV_d Q = Q 1 V 1 … Q d V d consists of non-empty and pairwise disjoint sets of variables V 1 , … , V d V_1,\ldots,V_d V 1 , … , V d (the quantifier blocks ) and quantifiers Q 1 , … , Q d ∈ { ∃ , ∀ } Q_1,\ldots,Q_d \in\{\exists,\forall\} Q 1 , … , Q d ∈ { ∃ , ∀ } with Q i ≠ Q i + 1 Q_i \ne Q_{i+1} Q i = Q i + 1 for i < d i < d i < d . A QBF is propositional if its prefix is empty or consists of a single existential block. In this case, the QBF is true if and only if its matrix is satisfiable, and we identify the QBF with its matrix.
A Q C N F \mathrm{QCNF} QCNF is a QBF Φ = Q ⋅ φ \Phi= Q \cdot\varphi Φ = Q ⋅ φ with C N F \mathrm{CNF} CNF φ \varphi φ . We define var ( Φ ) : = ⋃ C ∈ Φ var ( C ) \operatorname{var}(\Phi) := \bigcup_{C \in\Phi} \operatorname{var}(C) var ( Φ ) := ⋃ C ∈ Φ var ( C ) . We often identify a C N F \mathrm{CNF} CNF with the set of its clauses and write φ ⊆ ψ \varphi\subseteq\psi φ ⊆ ψ if every clause of φ \varphi φ is a clause of ψ \psi ψ . Analogously, for Q C N F s \mathrm{QCNFs} QCNFs Φ = Q ⋅ φ \Phi= Q \cdot\varphi Φ = Q ⋅ φ and Ψ = Q ′ ⋅ ψ \Psi= Q' \cdot\psi Ψ = Q ′ ⋅ ψ with the same prefix we write Φ ⊆ Ψ \Phi\subseteq\Psi Φ ⊆ Ψ if φ ⊆ ψ \varphi\subseteq\psi φ ⊆ ψ . We restrict a C N F \mathrm{CNF} CNF φ \varphi φ by an assignment σ \sigma σ as φ ∣ σ : = ⋀ C ∈ φ non-satisfied C ∣ σ \varphi|_{\sigma} := \bigwedge_{C \in\varphi\text{ non-satisfied}} C|_{\sigma} φ ∣ σ := ⋀ C ∈ φ non-satisfied C ∣ σ and a Q C N F \mathrm{QCNF} QCNF Φ = Q ⋅ φ \Phi= Q \cdot\varphi Φ = Q ⋅ φ as Φ ∣ σ : = Q ∣ σ ⋅ φ ∣ σ \Phi|_{\sigma} := Q|_{\sigma} \cdot\varphi|_{\sigma} Φ ∣ σ := Q ∣ σ ⋅ φ ∣ σ , where Q ∣ σ Q|_{\sigma} Q ∣ σ is obtained by deleting all variables from Q Q Q that appear in σ \sigma σ . Note that variables that no longer occur in φ ∣ σ \varphi|_{\sigma} φ ∣ σ remain in the prefix – such variables are simply irrelevant for the truth value. Hence, restricting a Q C N F \mathrm{QCNF} QCNF by an assignment of a proper subset of its first block never deletes a quantifier block.
Throughout the paper we assume that the prefix of a Q C N F \mathrm{QCNF} QCNF ends with an existential block. This is no restriction, because the literals of a universal block at the end of the prefix can be removed from all clauses without changing the truth value (this is the universal reduction rule recalled below), after which the block can be dropped. Hence, we usually write the prefix of a Q C N F \mathrm{QCNF} QCNF as ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 \forall U_0\exists X_1\forall U_1 \ldots\exists X_m\forall U_m\exists X_{m+1} ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 , where only the block U 0 U_0 U 0 is allowed to be empty. In this notation, m m m is the number of universal blocks that are preceded by an existential block, that is, the number of alternations from ∃ \exists ∃ to ∀ \forall ∀ .
Let Φ \Phi Φ be a QCNF with the prefix ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 \forall U_0\exists X_1\forall U_1\ldots\exists X_m\forall U_m\exists X_{m+1} ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 . A (universal) strategy is a tuple of functions s = ( s U 0 , … , s U m ) s=(s_{U_0},\ldots,s_{U_m}) s = ( s U 0 , … , s U m ) with s U j : ⟨ X 1 ∪ … ∪ X j ⟩ ⟶ ⟨ U j ⟩ s_{U_j}:\langle X_1\cup\ldots\cup X_j\rangle\longrightarrow\langle U_j\rangle s U j : ⟨ X 1 ∪ … ∪ X j ⟩ ⟶ ⟨ U j ⟩ for j = 0 , … , m j=0,\ldots,m j = 0 , … , m . If for each γ = ( γ 1 , … , γ m + 1 ) ∈ ⟨ X 1 ∪ … ∪ X m + 1 ⟩ \gamma=(\gamma_1,\ldots,\gamma_{m+1})\in\langle X_1\cup\ldots\cup X_{m+1}\rangle γ = ( γ 1 , … , γ m + 1 ) ∈ ⟨ X 1 ∪ … ∪ X m + 1 ⟩ the assignment γ ∪ ⋃ j = 0 m s U j ( ⋃ i = 1 j γ i ) \gamma\cup\bigcup_{j=0}^{m}s_{U_j}\left(\bigcup_{i=1}^{j}\gamma_i\right) γ ∪ ⋃ j = 0 m s U j ( ⋃ i = 1 j γ i ) falsifies the matrix of Φ \Phi Φ , then s s s is called a universal winning strategy. We define the size of a universal (winning) strategy as ∣ s ∣ : = ∑ j = 0 m ∣ Image ( s U j ) ∣ |s|:=\sum_{j=0}^{m}|\operatorname{Image}(s_{U_j})| ∣ s ∣ := ∑ j = 0 m ∣ Image ( s U j ) ∣ .
Dually, an (existential) strategy is a tuple of functions s = ( s X 1 , … , s X m + 1 ) s=(s_{X_1},\ldots,s_{X_{m+1}}) s = ( s X 1 , … , s X m + 1 ) with s X j : ⟨ U 0 ∪ … ∪ U j − 1 ⟩ ⟶ ⟨ X j ⟩ s_{X_j}:\langle U_0\cup\ldots\cup U_{j-1}\rangle\longrightarrow\langle X_j\rangle s X j : ⟨ U 0 ∪ … ∪ U j − 1 ⟩ ⟶ ⟨ X j ⟩ for j = 1 , … , m + 1 j=1,\ldots,m+1 j = 1 , … , m + 1 . If for each γ = ( γ 0 , … , γ m ) ∈ ⟨ U 0 ∪ … ∪ U m ⟩ \gamma=(\gamma_0,\ldots,\gamma_m)\in\langle U_0\cup\ldots\cup U_m\rangle γ = ( γ 0 , … , γ m ) ∈ ⟨ U 0 ∪ … ∪ U m ⟩ the assignment γ ∪ ⋃ j = 1 m + 1 s X j ( ⋃ i = 0 j − 1 γ i ) \gamma\cup\bigcup_{j=1}^{m+1}s_{X_j}\left(\bigcup_{i=0}^{j-1}\gamma_i\right) γ ∪ ⋃ j = 1 m + 1 s X j ( ⋃ i = 0 j − 1 γ i ) satisfies the matrix of Φ \Phi Φ , then s s s is called an existential winning strategy. The size of an existential (winning) strategy is defined as ∣ s ∣ : = ∑ j = 1 m + 1 ∣ Image ( s X j ) ∣ |s|:=\sum_{j=1}^{m+1}|\operatorname{Image}(s_{X_j})| ∣ s ∣ := ∑ j = 1 m + 1 ∣ Image ( s X j ) ∣ . We will simply write s ( γ ) s(\gamma) s ( γ ) instead of ⋃ j = 0 m s U j ( ⋃ i = 1 j γ i ) \bigcup_{j=0}^{m}s_{U_j}\left(\bigcup_{i=1}^{j}\gamma_i\right) ⋃ j = 0 m s U j ( ⋃ i = 1 j γ i ) or ⋃ j = 1 m + 1 s X j ( ⋃ i = 0 j − 1 γ i ) \bigcup_{j=1}^{m+1}s_{X_j}\left(\bigcup_{i=0}^{j-1}\gamma_i\right) ⋃ j = 1 m + 1 s X j ( ⋃ i = 0 j − 1 γ i ) , respectively. It is well known that a QCNF is false if and only if it has a universal winning strategy, and true if and only if it has an existential winning strategy.
Often we need to follow the two-player game along a strategy. For this purpose we introduce the following notation.
Definition 1 (game notation). Let s s s be a universal strategy for a QCNF with the prefix ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 \forall U_0\exists X_1\forall U_1\ldots\exists X_m\forall U_m\exists X_{m+1} ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 , where U 0 U_0 U 0 is allowed to be empty, and let γ = ( γ 1 , … , γ j ) ∈ ⟨ X 1 ∪ … ∪ X j ⟩ \gamma=(\gamma_1,\ldots,\gamma_j)\in\langle X_1\cup\ldots\cup X_j\rangle γ = ( γ 1 , … , γ j ) ∈ ⟨ X 1 ∪ … ∪ X j ⟩ (γ \gamma γ may also be an assignment of a superset of these variables). For i ≤ j i\le j i ≤ j we define
game ( s , γ , X i ) : = γ 1 ∪ … ∪ γ i ∪ ⋃ z = 0 i − 1 s U z ( γ 1 ∪ … ∪ γ z ) \operatorname{game}(s,\gamma,X_i):=\gamma_1\cup\ldots\cup\gamma_i\cup\bigcup_{z=0}^{i-1}s_{U_z}(\gamma_1\cup\ldots\cup\gamma_z) game ( s , γ , X i ) := γ 1 ∪ … ∪ γ i ∪ z = 0 ⋃ i − 1 s U z ( γ 1 ∪ … ∪ γ z ) and
game ( s , γ , U i ) : = γ 1 ∪ … ∪ γ i ∪ ⋃ z = 0 i s U z ( γ 1 ∪ … ∪ γ z ) . \operatorname{game}(s,\gamma,U_i):=\gamma_1\cup\ldots\cup\gamma_i\cup\bigcup_{z=0}^{i}s_{U_z}(\gamma_1\cup\ldots\cup\gamma_z). game ( s , γ , U i ) := γ 1 ∪ … ∪ γ i ∪ z = 0 ⋃ i s U z ( γ 1 ∪ … ∪ γ z ) . In other words , game ( s , γ , X i ) ∈ ⟨ U 0 ∪ X 1 ∪ … ∪ U i − 1 ∪ X i ⟩ \operatorname{game}(s,\gamma,X_i)\in\langle U_0\cup X_1\cup\ldots\cup U_{i-1}\cup X_i\rangle game ( s , γ , X i ) ∈ ⟨ U 0 ∪ X 1 ∪ … ∪ U i − 1 ∪ X i ⟩ and game ( s , γ , U i ) ∈ ⟨ U 0 ∪ X 1 ∪ … ∪ X i ∪ U i ⟩ \operatorname{game}(s,\gamma,U_i)\in\langle U_0\cup X_1\cup\ldots\cup X_i\cup U_i\rangle game ( s , γ , U i ) ∈ ⟨ U 0 ∪ X 1 ∪ … ∪ X i ∪ U i ⟩ are the assignments that are created during the two-player game up to the block ∃ X i \exists X_i ∃ X i (resp. ∀ U i \forall U_i ∀ U i ) if we follow the strategy s s s and the assignment γ \gamma γ for the existential variables. For convenience, we also set game ( s , γ , X 0 ) : = ∅ \operatorname{game}(s,\gamma,X_0):=\emptyset game ( s , γ , X 0 ) := ∅ , so that Φ ∣ game ( s , γ , X 0 ) = Φ \Phi|_{\operatorname{game}(s,\gamma,X_0)}=\Phi Φ ∣ game ( s , γ , X 0 ) = Φ . Analogously, for an existential strategy s s s and γ = ( γ 0 , … , γ j ) ∈ ⟨ U 0 ∪ … ∪ U j ⟩ \gamma=(\gamma_0,\ldots,\gamma_j)\in\langle U_0\cup\ldots\cup U_j\rangle γ = ( γ 0 , … , γ j ) ∈ ⟨ U 0 ∪ … ∪ U j ⟩ we define game ( s , γ , U i ) : = γ 0 ∪ … ∪ γ i ∪ ⋃ z = 1 i s X z ( γ 0 ∪ … ∪ γ z − 1 ) \operatorname{game}(s,\gamma,U_i):=\gamma_0\cup\ldots\cup\gamma_i\cup\bigcup_{z=1}^{i}s_{X_z}(\gamma_0\cup\ldots\cup\gamma_{z-1}) game ( s , γ , U i ) := γ 0 ∪ … ∪ γ i ∪ ⋃ z = 1 i s X z ( γ 0 ∪ … ∪ γ z − 1 ) and game ( s , γ , X i ) : = γ 0 ∪ … ∪ γ i − 1 ∪ ⋃ z = 1 i s X z ( γ 0 ∪ … ∪ γ z − 1 ) \operatorname{game}(s,\gamma,X_i):=\gamma_0\cup\ldots\cup\gamma_{i-1}\cup\bigcup_{z=1}^{i}s_{X_z}(\gamma_0\cup\ldots\cup\gamma_{z-1}) game ( s , γ , X i ) := γ 0 ∪ … ∪ γ i − 1 ∪ ⋃ z = 1 i s X z ( γ 0 ∪ … ∪ γ z − 1 ) for i ≤ j i\le j i ≤ j (resp. i ≤ j + 1 i\le j+1 i ≤ j + 1 ).
A proof system (cf. [22 ] ) for a language L L L is a poly-time computable function P : { 0 , 1 } ∗ → { 0 , 1 } ∗ P:\{0,1\}^{*}\to\{0,1\}^{*} P : { 0 , 1 } ∗ → { 0 , 1 } ∗ with Image ( P ) = L \operatorname{Image}(P)=L Image ( P ) = L . A P P P -proof π \pi π of some Φ ∈ L \Phi\in L Φ ∈ L is an element of { 0 , 1 } ∗ \{0,1\}^{*} { 0 , 1 } ∗ such that P ( π ) = Φ P(\pi)=\Phi P ( π ) = Φ . The inclusion Image ( P ) ⊆ L \operatorname{Image}(P)\subseteq L Image ( P ) ⊆ L is called soundness, while Image ( P ) ⊇ L \operatorname{Image}(P)\supseteq L Image ( P ) ⊇ L is the completeness of P P P . Here we consider the languages of unsatisfiable CNFs, false QCNFs and true QCNFs. A proof system P 1 P_1 P 1 simulates a proof system P 2 P_2 P 2 if there exists a polynomial p p p such that for each P 2 P_2 P 2 -proof π \pi π of a formula Φ \Phi Φ there is a P 1 P_1 P 1 -proof π ′ \pi' π ′ of Φ \Phi Φ with ∣ π ′ ∣ ≤ p ( ∣ π ∣ + ∣ Φ ∣ ) |\pi'|\le p(|\pi|+|\Phi|) ∣ π ′ ∣ ≤ p ( ∣ π ∣ + ∣Φ∣ ) . Here ∣ Φ ∣ |\Phi| ∣Φ∣ is the size of the formula. We include it because we measure the size of proofs by the number of their lines or steps, and a proof may be much smaller than the formula it proves. We say that P 1 P_1 P 1 p p p -simulates P 2 P_2 P 2 (cf. [22 ] ) if π ′ \pi' π ′ can even be computed from π \pi π by a poly-time computable function f f f . In this case we have P 1 ( f ( π ) ) = P 2 ( π ) P_1(f(\pi))=P_2(\pi) P 1 ( f ( π )) = P 2 ( π ) . If P 1 P_1 P 1 and P 2 P_2 P 2 p-simulate each other, we write P 1 ≡ p P 2 P_1\equiv_p P_2 P 1 ≡ p P 2 . Two proof systems are incomparable if neither simulates the other. We call P 2 P_2 P 2 exponentially weaker than P 1 P_1 P 1 if P 1 P_1 P 1 simulates P 2 P_2 P 2 and there is a family of formulas that requires P 2 P_2 P 2 -proofs of exponential size, while it has P 1 P_1 P 1 -proofs of polynomial size. We also use these notions relative to a class of formulas, for instance for QCNFs of bounded quantifier alternation.
We often add an N P NP N P oracle to a proof system. While the details depend on the proof system, the idea is that P proofs can evaluate SAT calls for free [12 ] . For such systems, the size of a proof counts a SAT call as a single step, and the propositional refutations that justify the calls are not counted (the proofs stay polynomially verifiable if these refutations are included). All simulations and separations in this paper that involve such a system refer to this measure.
Resolution (Res) is a proof system for unsatisfiable CNFs. A Res proof (or refutation ) π \pi π of a CNF φ \varphi φ is a sequence of clauses π = ( C 1 , … , C k ) \pi= (C_1,\ldots,C_k) π = ( C 1 , … , C k ) , where each C i C_i C i is either an axiom C i ∈ φ C_i \in\varphi C i ∈ φ or derived by the Resolution rule C i = ( C a ∨ C b ) ∖ { x , x ˉ } C_i = (C_a \vee C_b) \setminus\{x,\bar{x}\} C i = ( C a ∨ C b ) ∖ { x , x ˉ } with x ∈ C a x \in C_a x ∈ C a , x ˉ ∈ C b \bar{x} \in C_b x ˉ ∈ C b and a , b < i a,b < i a , b < i , and where C k = ( ⊥ ) C_k = (\bot) C k = ( ⊥ ) is the empty clause. We define the size ∣ π ∣ |\pi| ∣ π ∣ of π \pi π as the number k k k of its clauses.
Res can be lifted to a proof system Q-Resolution (Q-Res) [27 ] for false QCNFs by restricting the Resolution rule to existentially quantified literals x x x , disallowing the derivation of tautological clauses, and adding the Reduction rule C i = C j ∖ { u 1 , … , u n } C_i = C_j \setminus\{u_1,\ldots,u_n\} C i = C j ∖ { u 1 , … , u n } , where u 1 , … , u n u_1,\ldots,u_n u 1 , … , u n are universal literals such that there are no existential literals in C j C_j C j that are quantified right of each u h u_h u h . As for Res, a Q-Res refutation ends with the empty clause.
We can add an NP-oracle to Q-Res and obtain QN P {}^{\mathrm{NP}} NP -Res [5 ] by replacing the Resolution rule with the following Oracle rule , which derives a non-tautological clause C i C_i C i such that for some G ⊆ { C 1 , … , C i − 1 } \mathcal{G} \subseteq\{C_1,\ldots,C_{i-1}\} G ⊆ { C 1 , … , C i − 1 } :
⋀ C ∈ G C ∃ ⊨ C i ∃ \bigwedge_{C\in\mathcal{G}} C^{\exists} \models C_i^{\exists} C ∈ G ⋀ C ∃ ⊨ C i ∃
and C i ∀ C_i^{\forall} C i ∀ is a subclause of C ∀ C^{\forall} C ∀ for each C ∈ G C \in\mathcal{G} C ∈ G . Here C ∃ C^{\exists} C ∃ (resp. C ∀ C^{\forall} C ∀ ) denotes the existential (resp. universal) subclause of C C C .
CAQE: the basic algorithmic paradigm
CAQE, which stands for Clausal Abstraction for Quantifier Elimination , was introduced in 2015 [35 ] . It is a recursive algorithm that works on the first quantifier block of a QCNF and eliminates the quantifiers one by one. The idea is simple, but elegant: Suppose we have a QCNF Φ = ∃ V Q ⋅ φ \Phi= \exists V Q \cdot\varphi Φ = ∃ V Q ⋅ φ , such that ∃ V \exists V ∃ V is the first quantifier block and Q Q Q are the remaining quantifiers. Our goal is to find out if Φ \Phi Φ is true or false. Obviously, Φ \Phi Φ is true if and only if there exists an assignment α ∈ ⟨ V ⟩ \alpha\in\langle V\rangle α ∈ ⟨ V ⟩ such that Φ ∣ α = Q ⋅ φ ∣ α \Phi|_\alpha= Q \cdot\varphi|_\alpha Φ ∣ α = Q ⋅ φ ∣ α is true, whereas the truth value of Φ ∣ α \Phi|_\alpha Φ ∣ α can be checked recursively. If we can find at least one such α \alpha α , we are done and we have proven the truth of Φ \Phi Φ .
However, if Φ \Phi Φ is a false formula, then we need to check that for all assignments α ∈ ⟨ V ⟩ \alpha\in\langle V\rangle α ∈ ⟨ V ⟩ the formula Φ ∣ α \Phi|_\alpha Φ ∣ α is indeed false. If we go through all of these assignments, we end up with a runtime exponential in ∣ V ∣ |V| ∣ V ∣ . The trick is now to pick specific assignments α \alpha α that suffice to show the falsity of Φ \Phi Φ . More precisely, we want to pick the “optimal” or “best” assignments α \alpha α which satisfy the “highest number of” clauses from Φ \Phi Φ . If these optimal assignments do not succeed in making Φ \Phi Φ true, then neither will the assignments that are “worse”. Let us demonstrate this on an example:
Example 2. Let Φ \Phi Φ be the QCNF with prefix ∃ x 1 , x 2 ∀ u 1 , u 2 ∃ y 1 , y 2 \exists x_1,x_2\forall u_1,u_2\exists y_1,y_2 ∃ x 1 , x 2 ∀ u 1 , u 2 ∃ y 1 , y 2 and the matrix
x 1 ∨ u 1 ∨ y 1 x 1 ∨ x 2 ∨ u 2 ∨ y 2 y ˉ 1 ∨ y ˉ 2 x ˉ 1 ∨ u ˉ 1 ∨ y 1 x ˉ 1 ∨ u ˉ 2 ∨ y 2 x ˉ 2 ∨ u ˉ 2 ∨ y 2 . \begin{aligned}
&x_1 \vee u_1 \vee y_1 \qquad x_1 \vee x_2 \vee u_2 \vee y_2 \qquad\bar{y}_1 \vee\bar{y}_2 \\
&\bar{x}_1 \vee\bar{u}_1 \vee y_1 \qquad\bar{x}_1 \vee\bar{u}_2 \vee y_2 \\
&\bar{x}_2 \vee\bar{u}_2 \vee y_2.
\end{aligned} x 1 ∨ u 1 ∨ y 1 x 1 ∨ x 2 ∨ u 2 ∨ y 2 y ˉ 1 ∨ y ˉ 2 x ˉ 1 ∨ u ˉ 1 ∨ y 1 x ˉ 1 ∨ u ˉ 2 ∨ y 2 x ˉ 2 ∨ u ˉ 2 ∨ y 2 .
This formula is false since the universal player wins by setting u 1 ≡ x 1 u_1 \equiv x_1 u 1 ≡ x 1 and u 2 ≡ x 1 ∨ x 2 u_2 \equiv x_1 \vee x_2 u 2 ≡ x 1 ∨ x 2 . Now, there are four assignments for { x 1 , x 2 } \{x_1,x_2\} { x 1 , x 2 } that would need to be tested: α 1 = ( x 1 ↦ 0 , x 2 ↦ 0 ) \alpha_1 = (x_1 \mapsto0,x_2 \mapsto0) α 1 = ( x 1 ↦ 0 , x 2 ↦ 0 ) , α 2 = ( x 1 ↦ 0 , x 2 ↦ 1 ) \alpha_2 = (x_1 \mapsto0,x_2 \mapsto1) α 2 = ( x 1 ↦ 0 , x 2 ↦ 1 ) , α 3 = ( x 1 ↦ 1 , x 2 ↦ 0 ) \alpha_3 = (x_1 \mapsto1,x_2 \mapsto0) α 3 = ( x 1 ↦ 1 , x 2 ↦ 0 ) and α 4 = ( x 1 ↦ 1 , x 2 ↦ 1 ) \alpha_4 = (x_1 \mapsto1,x_2 \mapsto1) α 4 = ( x 1 ↦ 1 , x 2 ↦ 1 ) .
However, we can observe that α 3 \alpha_3 α 3 already satisfies all the clauses that are satisfied by α 4 \alpha_4 α 4 . Even worse, α 4 \alpha_4 α 4 satisfies fewer clauses than α 3 \alpha_3 α 3 . Hence, if we have shown that Φ ∣ α 3 \Phi|_{\alpha_3} Φ ∣ α 3 is false, this also implies that Φ ∣ α 4 \Phi|_{\alpha_4} Φ ∣ α 4 must be false as well. Therefore, if we test α 3 \alpha_3 α 3 before α 4 \alpha_4 α 4 , we can omit testing α 4 \alpha_4 α 4 altogether. In other words, checking α 1 \alpha_1 α 1 , α 2 \alpha_2 α 2 , α 3 \alpha_3 α 3 suffices to prove the falsity of Φ \Phi Φ .
An important aspect of CAQE is the direct inclusion of a SAT solver. Essentially, the task of the SAT solver, usually implemented as CDCL, is to return an assignment α j ∈ ⟨ V ⟩ \alpha_j \in\langle V\rangle α j ∈ ⟨ V ⟩ such that α j \alpha_j α j is not “worse” than or equal to one of the previously tested α 1 , … , α j − 1 \alpha_1,\ldots,\alpha_{j-1} α 1 , … , α j − 1 . Or more precisely, the SAT solver will return an α j \alpha_j α j such that for each i = 1 , … , j − 1 i=1,\ldots,j-1 i = 1 , … , j − 1 the assignment α j \alpha_j α j satisfies at least one clause that is not satisfied by α i \alpha_i α i . That also means that if the SAT solver cannot find such an α j \alpha_j α j we can conclude that the formula is false.
The procedure for formulas Φ = ∀ V Q ⋅ φ \Phi= \forall V Q \cdot\varphi Φ = ∀ V Q ⋅ φ that start with a universal quantifier ∀ V \forall V ∀ V is similar. The main difference is that the roles of universal and existential blocks are swapped and “best” or optimal assignments α ∈ ⟨ V ⟩ \alpha\in\langle V\rangle α ∈ ⟨ V ⟩ are assignments that satisfy as few clauses as possible.
In the next section we will explain in further detail how CAQE compares the α \alpha α in terms of optimality. A high-level overview of the functionality of CAQE is given in Figure 2 .
Figure 2. High-level flowchart of the basic CAQE algorithm. Yellow nodes are performed by the SAT solver. Blue nodes are performed recursively.
Formalising basic CAQE as a proof system
Intuitively, each clause C r C_r C r of the given QCNF Φ \Phi Φ is split into two parts: the top part and the bottom part. The top part, which we denote as C r , 1 C_{r,1} C r , 1 , consists of variables from the first quantifier block and contains all the information that we need to find all optimal assignments of that block. The bottom part, denoted as C r , > 1 C_{r,>1} C r , > 1 , contains the rest of the clause and is needed for the recursive call. Furthermore, C r , 1 C_{r,1} C r , 1 and C r , > 1 C_{r,>1} C r , > 1 are connected by an auxiliary variable b r b_r b r (sometimes also called a bottom variable ) which we need to encode whether or not an assignment has satisfied C r C_r C r on the first block. For each clause C r C_r C r we have a separate variable b r b_r b r , and we define B B B as the set of all auxiliary variables.
Definition 3. Let Φ : = Q V 1 Q ⋅ C 1 ∧ … ∧ C k \Phi:= QV_1Q \cdot C_1 \land\ldots\land C_k Φ := Q V 1 Q ⋅ C 1 ∧ … ∧ C k with clauses C 1 , … , C k C_1,\ldots,C_k C 1 , … , C k . We define auxiliary variables B : = { b 1 , … , b k } B := \{b_1,\ldots,b_k\} B := { b 1 , … , b k } with B ∩ var ( Φ ) = ∅ B \cap\operatorname{var}(\Phi) = \varnothing B ∩ var ( Φ ) = ∅ . If Q = ∃ Q = \exists Q = ∃ , we define
C r , 1 : = ⋁ ℓ ∈ C r var ( ℓ ) ∈ V 1 ℓ ∨ b r and C r , > 1 : = ⋁ ℓ ∈ C r var ( ℓ ) ∉ V 1 ℓ ∨ b r ‾ . C_{r,1} := \bigvee_{\substack{\ell\in C_r\\ \operatorname{var}(\ell) \in V_1}} \ell\lor b_r
\qquad\text{and}\qquad
C_{r,>1} := \bigvee_{\substack{\ell\in C_r\\ \operatorname{var}(\ell) \notin V_1}} \ell\lor\overline{b_r}. C r , 1 := ℓ ∈ C r var ( ℓ ) ∈ V 1 ⋁ ℓ ∨ b r and C r , > 1 := ℓ ∈ C r var ( ℓ ) ∈ / V 1 ⋁ ℓ ∨ b r . We set
Φ ∃ V 1 : = ⋀ r = 1 k C r , 1 and Φ > 1 : = Q ⋅ ⋀ r = 1 k C r , > 1 . \Phi_{\exists V_1} := \bigwedge_{r=1}^{k} C_{r,1}
\qquad\text{and}\qquad
\Phi_{>1} := Q \cdot\bigwedge_{r=1}^{k} C_{r,>1}. Φ ∃ V 1 := r = 1 ⋀ k C r , 1 and Φ > 1 := Q ⋅ r = 1 ⋀ k C r , > 1 . Dually, if Q = ∀ Q = \forall Q = ∀ , we define
C r , 1 : = ⋀ ℓ ∈ C r var ( ℓ ) ∈ V 1 ( ℓ ‾ ∨ b r ‾ ) and C r , > 1 : = ⋁ ℓ ∈ C r var ( ℓ ) ∉ V 1 ℓ ∨ b r ‾ C_{r,1} := \bigwedge_{\substack{\ell\in C_r\\ \operatorname{var}(\ell) \in V_1}} \left(\overline{\ell} \lor\overline{b_r}\right)
\qquad\text{and}\qquad
C_{r,>1} := \bigvee_{\substack{\ell\in C_r\\ \operatorname{var}(\ell) \notin V_1}} \ell\lor\overline{b_r} C r , 1 := ℓ ∈ C r var ( ℓ ) ∈ V 1 ⋀ ( ℓ ∨ b r ) and C r , > 1 := ℓ ∈ C r var ( ℓ ) ∈ / V 1 ⋁ ℓ ∨ b r and set
Φ ∀ V 1 : = ⋀ r = 1 k C r , 1 and Φ > 1 : = Q ⋅ ⋀ r = 1 k C r , > 1 . \Phi_{\forall V_1} := \bigwedge_{r=1}^{k} C_{r,1}
\qquad\text{and}\qquad
\Phi_{>1} := Q \cdot\bigwedge_{r=1}^{k} C_{r,>1}. Φ ∀ V 1 := r = 1 ⋀ k C r , 1 and Φ > 1 := Q ⋅ r = 1 ⋀ k C r , > 1 . For an assignment β ∈ ⟨ B ⟩ \beta\in\langle B\rangle β ∈ ⟨ B ⟩ , we define
Φ > 1 ∣ β : = Q ⋅ ⋀ 1 ≤ r ≤ k C r , > 1 ∣ β = Q ⋅ ⋀ 1 ≤ r ≤ k β ( b r ) = 1 C r , > 1 ∣ β . \Phi_{>1\mid\beta} := Q \cdot\bigwedge_{1\leq r\leq k} C_{r,>1\mid\beta}
= Q \cdot\bigwedge_{\substack{1\leq r\leq k\\ \beta(b_r)=1}} C_{r,>1\mid\beta}. Φ > 1 ∣ β := Q ⋅ 1 ≤ r ≤ k ⋀ C r , > 1 ∣ β = Q ⋅ 1 ≤ r ≤ k β ( b r ) = 1 ⋀ C r , > 1 ∣ β . Note that Φ ∃ V 1 \Phi_{\exists V_1} Φ ∃ V 1 and Φ ∀ V 1 \Phi_{\forall V_1} Φ ∀ V 1 are propositional formulas, while both Φ > 1 ∣ β \Phi_{>1\mid\beta} Φ > 1 ∣ β and Φ > 1 \Phi_{>1} Φ > 1 are QCNFs.1
In the whole paper, clause indices refer to the clauses C 1 , … , C k C_1,\ldots,C_k C 1 , … , C k of the QCNF Φ \Phi Φ we are currently considering. All formulas that we obtain from Φ \Phi Φ by restrictions and by the splitting of Definition 3 consist of subclauses of the clauses C r C_r C r . Such a subclause keeps the index r r r of the clause it comes from. In the same way, the auxiliary variables B = { b 1 , … , b k } B = \{b_1,\ldots,b_k\} B = { b 1 , … , b k } are the same on all levels of the recursion. Hence, we can speak of “the clause C r C_r C r ” of a restricted formula, and we can say that the clause C r C_r C r is present in or satisfied in a restricted formula. Accordingly, we regard a restricted formula as its prefix together with the set of its indexed clauses: two formulas obtained from Φ \Phi Φ by restrictions are equal if they have the same prefix and, for every index r r r , either both do not contain C r C_r C r or both contain the same subclause of C r C_r C r under the index r r r . For a restricted formula whose clauses have the index set J ⊆ { 1 , … , k } J \subseteq\{1,\ldots,k\} J ⊆ { 1 , … , k } , the auxiliary variables are the b r b_r b r with r ∈ J r \in J r ∈ J , and a canonical assignment (Definition 4 below) only assigns these variables.
The auxiliary B B B -variables that are assigned by β \beta β indicate whether a clause is already satisfied in the first quantifier block, or if it might need to be satisfied using the remaining blocks. The assignment b r ↦ 1 b_r \mapsto1 b r ↦ 1 encodes the information “C r C_r C r might need to be satisfied in the bottom part”, while b r ↦ 0 b_r \mapsto0 b r ↦ 0 means “C r C_r C r is already satisfied and does not need to be satisfied by the bottom part”. More precisely, the propositional formula Φ ∃ V 1 \Phi_{\exists V_1} Φ ∃ V 1 (resp. Φ ∀ V 1 \Phi_{\forall V_1} Φ ∀ V 1 ) is given to a SAT solver which returns a satisfying assignment α ∪ β ∈ ⟨ V 1 ∪ B ⟩ \alpha\cup\beta\in\langle V_1 \cup B\rangle α ∪ β ∈ ⟨ V 1 ∪ B ⟩ . Such an assignment always exists (cf. Lemma 5 below). In the case ∃ V 1 \exists V_1 ∃ V 1 , the variable b r b_r b r must be set to 1 if α \alpha α does not satisfy C r C_r C r , otherwise it is allowed (but not forced) to be set to 0.
Those clauses C r C_r C r for which b r b_r b r is set to 1 are precisely the clauses that remain in Φ > 1 \Phi_{>1} Φ > 1 , hence Φ > 1 ∣ β \Phi_{>1}\mathbin{|}\beta Φ > 1 ∣ β contains at least all the clauses that are not satisfied by the corresponding α \alpha α . That means if Φ > 1 ∣ β \Phi_{>1}\mathbin{|}\beta Φ > 1 ∣ β is true, Φ \Phi Φ turns out to be true as well.
In the case ∀ V 1 \forall V_1 ∀ V 1 , the b r b_r b r are interpreted a bit differently. Here b r b_r b r needs to be set to 0 if α \alpha α satisfies C r C_r C r , and b r b_r b r is allowed (but not forced) to be set to 1 if α \alpha α does not satisfy C r C_r C r . Therefore all clauses that remain in Φ > 1 ∣ β \Phi_{>1}\mathbin{|}\beta Φ > 1 ∣ β are clauses that are not satisfied by α \alpha α . If Φ > 1 ∣ β \Phi_{>1}\mathbin{|}\beta Φ > 1 ∣ β turns out to be false, then this also shows the falsity of Φ \Phi Φ itself.
In both cases there is a special choice of β \beta β that sets exactly the auxiliary variables of the clauses not satisfied by α \alpha α to 1. In the existential case, this is the minimal satisfying assignment b min b_{\min} b m i n of Φ ∃ V 1 \Phi_{\exists V_1} Φ ∃ V 1 in the terminology of [35 ] , and in the universal case it is the maximal satisfying assignment b max b_{\max} b m a x of Φ ∀ V 1 \Phi_{\forall V_1} Φ ∀ V 1 .
Definition 4 (canonical assignments). Let Φ \Phi Φ be a Q C N F QCNF QC N F with clauses C 1 , … , C k C_1,\ldots,C_k C 1 , … , C k such that its prefix starts with the block Q V 1 QV_1 Q V 1 , and let α ∈ ⟨ V 1 ⟩ \alpha\in\langle V_1\rangle α ∈ ⟨ V 1 ⟩ . The assignment β ∈ ⟨ B ⟩ \beta\in\langle B\rangle β ∈ ⟨ B ⟩ is called canonical for α \alpha α if for all r = 1 , … , k r=1,\ldots,k r = 1 , … , k it holds that β ( b r ) = 0 \beta(b_r)=0 β ( b r ) = 0 if and only if α \alpha α satisfies C r C_r C r .
Lemma 5. Let Φ \Phi Φ be a Q C N F QCNF QC N F with the matrix C 1 ∧ … ∧ C k C_1 \land\ldots\land C_k C 1 ∧ … ∧ C k that starts with the quantifier block Q V 1 QV_1 Q V 1 with Q ∈ { ∃ , ∀ } Q \in\{\exists,\forall\} Q ∈ { ∃ , ∀ } , let α ∈ ⟨ V 1 ⟩ \alpha\in\langle V_1\rangle α ∈ ⟨ V 1 ⟩ be arbitrary and let β ∈ ⟨ B ⟩ \beta\in\langle B\rangle β ∈ ⟨ B ⟩ be canonical for α \alpha α . Then α ∪ β \alpha\cup\beta α ∪ β satisfies Φ Q V 1 \Phi_{QV_1} Φ Q V 1 .
Proof. Let us first consider the case Q = ∃ Q=\exists Q = ∃ . Then we have
Φ ∃ V 1 = ⋀ r = 1 k C r , 1 with C r , 1 = ⋁ ℓ ∈ C r , var ( ℓ ) ∈ V 1 ℓ ∨ b r for r = 1 , … , k . \Phi_{\exists V_1}=\bigwedge_{r=1}^{k} C_{r,1}\text{ with }C_{r,1}=\bigvee_{\ell\in C_r,\ \operatorname{var}(\ell)\in V_1}\ell\vee b_r\text{ for }r=1,\ldots,k. Φ ∃ V 1 = r = 1 ⋀ k C r , 1 with C r , 1 = ℓ ∈ C r , var ( ℓ ) ∈ V 1 ⋁ ℓ ∨ b r for r = 1 , … , k .
If α \alpha α satisfies C r C_r C r , then it does so by a literal from V 1 V_1 V 1 , hence it also satisfies C r , 1 C_{r,1} C r , 1 . If α \alpha α does not satisfy C r C_r C r , then we have β ( b r ) = 1 \beta(b_r)=1 β ( b r ) = 1 , therefore β \beta β satisfies C r , 1 C_{r,1} C r , 1 . Hence, α ∪ β \alpha\cup\beta α ∪ β satisfies all C r , 1 C_{r,1} C r , 1 for r = 1 , … , k r=1,\ldots,k r = 1 , … , k , and therefore also Φ Q V 1 = Φ ∃ V 1 \Phi_{QV_1}=\Phi_{\exists V_1} Φ Q V 1 = Φ ∃ V 1 .
In the case Q = ∀ Q=\forall Q = ∀ we have
Φ ∀ V 1 = ⋀ r = 1 k C r , 1 with C r , 1 = ⋁ ℓ ∈ C r , var ( ℓ ) ∈ V 1 ( ℓ ‾ ∨ b r ‾ ) for r = 1 , … , k . \Phi_{\forall V_1}=\bigwedge_{r=1}^{k} C_{r,1}\text{ with }C_{r,1}=\bigvee_{\ell\in C_r,\ \operatorname{var}(\ell)\in V_1}(\overline{\ell}\vee\overline{b_r})\text{ for }r=1,\ldots,k. Φ ∀ V 1 = r = 1 ⋀ k C r , 1 with C r , 1 = ℓ ∈ C r , var ( ℓ ) ∈ V 1 ⋁ ( ℓ ∨ b r ) for r = 1 , … , k .
If α \alpha α satisfies C r C_r C r , then we have β ( b r ) = 0 \beta(b_r)=0 β ( b r ) = 0 and therefore β \beta β satisfies C r , 1 C_{r,1} C r , 1 . If α \alpha α does not satisfy C r C_r C r , then for all ℓ ∈ C r \ell\in C_r ℓ ∈ C r with var ( ℓ ) ∈ V 1 \operatorname{var}(\ell)\in V_1 var ( ℓ ) ∈ V 1 we have α ( ℓ ) = 0 \alpha(\ell)=0 α ( ℓ ) = 0 . Hence, α \alpha α satisfies C r , 1 C_{r,1} C r , 1 . We conclude that α ∪ β \alpha\cup\beta α ∪ β satisfies all C r , 1 C_{r,1} C r , 1 for r = 1 , … , k r=1,\ldots,k r = 1 , … , k and therefore also Φ Q V 1 = Φ ∀ V 1 \Phi_{QV_1}=\Phi_{\forall V_1} Φ Q V 1 = Φ ∀ V 1 . ◀
Example 6. Let Φ \Phi Φ be the formula from Example 2. Then Φ ∃ x 1 , x 2 \Phi_{\exists x_1,x_2} Φ ∃ x 1 , x 2 consists of the clauses
x 1 ∨ b 1 x 1 ∨ x 2 ∨ b 2 ( b 3 ) x 1 ‾ ∨ b 4 x 1 ‾ ∨ b 5 x 2 ‾ ∨ b 6 x_1\vee b_1\qquad x_1\vee x_2\vee b_2\qquad(b_3)\qquad\overline{x_1}\vee b_4\qquad\overline{x_1}\vee b_5\qquad\overline{x_2}\vee b_6 x 1 ∨ b 1 x 1 ∨ x 2 ∨ b 2 ( b 3 ) x 1 ∨ b 4 x 1 ∨ b 5 x 2 ∨ b 6 and Φ > 1 \Phi_{>1} Φ > 1 is the Q C N F QCNF QC N F with the prefix ∀ u 1 , u 2 ∃ y 1 , y 2 \forall u_1,u_2\exists y_1,y_2 ∀ u 1 , u 2 ∃ y 1 , y 2 and the matrix
b 1 ‾ ∨ u 1 ∨ y 1 b 2 ‾ ∨ u 2 ∨ y 2 b 3 ‾ ∨ y 1 ‾ ∨ y 2 ‾ b 4 ‾ ∨ u 1 ‾ ∨ y 1 b 5 ‾ ∨ u 2 ‾ ∨ y 2 b 6 ‾ ∨ u 2 ‾ ∨ y 2 . \overline{b_1}\vee u_1\vee y_1\qquad\overline{b_2}\vee u_2\vee y_2\qquad\overline{b_3}\vee\overline{y_1}\vee\overline{y_2}\qquad\overline{b_4}\vee\overline{u_1}\vee y_1\qquad\overline{b_5}\vee\overline{u_2}\vee y_2\qquad\overline{b_6}\vee\overline{u_2}\vee y_2. b 1 ∨ u 1 ∨ y 1 b 2 ∨ u 2 ∨ y 2 b 3 ∨ y 1 ∨ y 2 b 4 ∨ u 1 ∨ y 1 b 5 ∨ u 2 ∨ y 2 b 6 ∨ u 2 ∨ y 2 . For α 3 = ( x 1 ↦ 1 , x 2 ↦ 0 ) \alpha_3=(x_1\mapsto1,x_2\mapsto0) α 3 = ( x 1 ↦ 1 , x 2 ↦ 0 ) , the canonical assignment is β = b 1 ‾ ∧ b 2 ‾ ∧ b 3 ∧ b 4 ∧ b 5 ∧ b 6 ‾ \beta=\overline{b_1}\land\overline{b_2}\land b_3\land b_4\land b_5\land\overline{b_6} β = b 1 ∧ b 2 ∧ b 3 ∧ b 4 ∧ b 5 ∧ b 6 . Then Φ > 1 ∣ β \Phi_{>1}\mathbin{|}\beta Φ > 1 ∣ β is the Q C N F QCNF QC N F ∀ u 1 , u 2 ∃ y 1 , y 2 ⋅ ( u 1 ‾ ∨ y 2 ‾ ) ∧ ( u 1 ‾ ∨ y 1 ) ∧ ( u 2 ‾ ∨ y 2 ) = Φ ∣ α 3 \forall u_1,u_2\exists y_1,y_2\cdot(\overline{u_1}\vee\overline{y_2})\land(\overline{u_1}\vee y_1)\land(\overline{u_2}\vee y_2)=\Phi\mathbin{|}\alpha_3 ∀ u 1 , u 2 ∃ y 1 , y 2 ⋅ ( u 1 ∨ y 2 ) ∧ ( u 1 ∨ y 1 ) ∧ ( u 2 ∨ y 2 ) = Φ ∣ α 3 .
So far we have only explained how to obtain a solution α ∪ β \alpha\cup\beta α ∪ β for the first iteration when no other α , β \alpha,\beta α , β have been tested yet. Remember that we only want to test α \alpha α that are not worse than previous ones. Because this concept of optimality does not primarily depend on α \alpha α syntactically, we will instead use the β \beta β for that as β \beta β describes the effect of α \alpha α on the clauses. In the ∃ V 1 \exists V_1 ∃ V 1 case, we want to make sure that for each previous α \alpha α the new α \alpha α satisfies at least one clause that the previous α \alpha α did not. In other words, the new β \beta β has to flip at least one 1 from each previous β \beta β to 0. If β 1 , … , β j − 1 \beta_1,\ldots,\beta_{j-1} β 1 , … , β j − 1 are the previously tested assignments and β j \beta_j β j is the new one, then β j \beta_j β j has to satisfy the formula ⋀ a = 1 j − 1 ( ⋁ b ∈ β a − 1 ( 1 ) b ‾ ) \bigwedge_{a=1}^{j-1}\left(\bigvee_{b\in\beta_a^{-1}(1)}\overline{b}\right) ⋀ a = 1 j − 1 ( ⋁ b ∈ β a − 1 ( 1 ) b ) which we will simply add to Φ ∃ V 1 \Phi_{\exists V_1} Φ ∃ V 1 by conjunction. Analogously, in the case ∀ V 1 \forall V_1 ∀ V 1 we require that the new β \beta β flips at least one 0 from each previous β \beta β to 1. This can be enforced by ⋀ a = 1 j − 1 ( ⋁ b ∈ β a − 1 ( 0 ) b ) \bigwedge_{a=1}^{j-1}\left(\bigvee_{b\in\beta_a^{-1}(0)}b\right) ⋀ a = 1 j − 1 ( ⋁ b ∈ β a − 1 ( 0 ) b ) which we add to Φ ∀ V 1 \Phi_{\forall V_1} Φ ∀ V 1 by conjunction.2 We call these additional clauses refinement clauses and introduce the notation
G β ∃ : = ⋁ b ∈ β − 1 ( 1 ) b ‾ and G β ∀ : = ⋁ b ∈ β − 1 ( 0 ) b G^{\exists}_{\beta}:=\bigvee_{b\in\beta^{-1}(1)}\overline{b}
\qquad\text{and}\qquad
G^{\forall}_{\beta}:=\bigvee_{b\in\beta^{-1}(0)}b G β ∃ := b ∈ β − 1 ( 1 ) ⋁ b and G β ∀ := b ∈ β − 1 ( 0 ) ⋁ b
for β ∈ ⟨ B ⟩ \beta\in\langle B\rangle β ∈ ⟨ B ⟩ . It is easy to see that the refinement clauses can be reformulated as follows.
Lemma 8. Let β 1 , … , β g ∈ ⟨ B ⟩ \beta_1,\ldots,\beta_g\in\langle B\rangle β 1 , … , β g ∈ ⟨ B ⟩ . Then
β ∈ ⟨ B ⟩ \beta\in\langle B\rangle β ∈ ⟨ B ⟩ satisfies ⋀ a = 1 g G β a ∃ \bigwedge_{a=1}^{g}G^{\exists}_{\beta_a} ⋀ a = 1 g G β a ∃ if and only if for each a = 1 , … , g a=1,\ldots,g a = 1 , … , g there exists some b ∈ B b\in B b ∈ B such that β ( b ) < β a ( b ) \beta(b)<\beta_a(b) β ( b ) < β a ( b ) , which means β − 1 ( 1 ) ⊈ β a − 1 ( 1 ) \beta^{-1}(1)\not\subseteq\beta_a^{-1}(1) β − 1 ( 1 ) ⊆ β a − 1 ( 1 ) ,
β ∈ ⟨ B ⟩ \beta\in\langle B\rangle β ∈ ⟨ B ⟩ satisfies ⋀ a = 1 g G β a ∀ \bigwedge_{a=1}^{g}G^{\forall}_{\beta_a} ⋀ a = 1 g G β a ∀ if and only if for each a = 1 , … , g a=1,\ldots,g a = 1 , … , g there exists some b ∈ B b\in B b ∈ B such that β ( b ) > β a ( b ) \beta(b)>\beta_a(b) β ( b ) > β a ( b ) , which means β − 1 ( 0 ) ⊈ β a − 1 ( 0 ) \beta^{-1}(0)\not\subseteq\beta_a^{-1}(0) β − 1 ( 0 ) ⊆ β a − 1 ( 0 ) .
Proof. The statements become trivial by reformulating them: the assignment β ∈ ⟨ B ⟩ \beta\in\langle B\rangle β ∈ ⟨ B ⟩ satisfies ⋀ a = 1 g ⋁ b ∈ β a − 1 ( 1 ) b ‾ \bigwedge_{a=1}^{g}\bigvee_{b\in\beta_a^{-1}(1)}\overline{b} ⋀ a = 1 g ⋁ b ∈ β a − 1 ( 1 ) b if and only if for each a = 1 , … , g a=1,\ldots,g a = 1 , … , g there exists some b ∈ B b\in B b ∈ B such that β ( b ) = 0 \beta(b)=0 β ( b ) = 0 and β a ( b ) = 1 \beta_a(b)=1 β a ( b ) = 1 . Analogously, β ∈ ⟨ B ⟩ \beta\in\langle B\rangle β ∈ ⟨ B ⟩ satisfies ⋀ a = 1 g ⋁ b ∈ β a − 1 ( 0 ) b \bigwedge_{a=1}^{g}\bigvee_{b\in\beta_a^{-1}(0)}b ⋀ a = 1 g ⋁ b ∈ β a − 1 ( 0 ) b if and only if for each a = 1 , … , g a=1,\ldots,g a = 1 , … , g there exists some b ∈ B b\in B b ∈ B such that β ( b ) = 1 \beta(b)=1 β ( b ) = 1 and β a ( b ) = 0 \beta_a(b)=0 β a ( b ) = 0 .
Our goal to design a proof system that characterises the basic CAQE algorithm amounts to defining a structure that stores all the relevant information for each recursive call of the algorithm. We call these structures slices . A slice contains the following information: The current formula for that call, all α \alpha α that were tested, all β \beta β that correspond to α \alpha α , and references to further recursive calls for Φ > 1 ∣ β \Phi_{>1\mid\beta} Φ > 1 ∣ β (i.e. references to further slices) and a Res refutation ρ \rho ρ in case the propositional formula given to the SAT solver turns out to be unsatisfiable at some point. There are two exceptions: If Φ \Phi Φ itself is an unsatisfiable propositional formula, a slice simply consists of a Res refutation3 , while for a satisfiable propositional formula a slice contains a satisfying assignment only. This is due to the fact that CAQE simply calls the SAT solver once only a propositional formula remains (i.e. when we reach the last quantifier block).
Each recursive call of CAQE either proves that its formula is false or that it is true. Therefore, we distinguish between refutation slices and verification slices . We define the two notions by mutual recursion, in the same way as the algorithm recurses. Let us first consider a slice for a formula that starts with an existential block. It is a refutation slice if all recursive calls returned “false” and the SAT solver finally reported that no further assignment needs to be tested. It is a verification slice if the last recursive call returned “true”. For universal blocks the roles of true and false are swapped. We summarise this formally as follows.
Definition 9 (slices). Let Φ \Phi Φ be a Q C N F QCNF QC N F with prefix Q 1 V 1 Q 2 V 2 … Q d V d Q_{1}V_{1}Q_{2}V_{2}\ldots Q_{d}V_{d} Q 1 V 1 Q 2 V 2 … Q d V d , where V i V_{i} V i are disjoint sets of variables, Q i ∈ { ∃ , ∀ } Q_{i} \in\{\exists,\forall\} Q i ∈ { ∃ , ∀ } and Q i ≠ Q i + 1 Q_{i} \ne Q_{i+1} Q i = Q i + 1 for all i i i , and with clauses C 1 , … , C k C_{1},\ldots,C_{k} C 1 , … , C k . A slice S S S for Φ \Phi Φ together with its size ∣ S ∣ \lvert S\rvert ∣ S ∣ is one of the following. Each slice is either a refutation slice or a verification slice.
S = ( Φ , π ) S = (\Phi,\pi) S = ( Φ , π ) , where Φ \Phi Φ is propositional and π \pi π is a Res refutation of Φ \Phi Φ ; ∣ S ∣ : = 1 \lvert S\rvert:= 1 ∣ S ∣ := 1 . S S S is a refutation slice.
S = ( Φ , σ ) S = (\Phi,\sigma) S = ( Φ , σ ) , where Φ \Phi Φ is propositional and σ ∈ ⟨ var ( Φ ) ⟩ \sigma\in\langle\operatorname{var}(\Phi)\rangle σ ∈ ⟨ var ( Φ )⟩ satisfies Φ \Phi Φ ; ∣ S ∣ : = 1 \lvert S\rvert:= 1 ∣ S ∣ := 1 . S S S is a verification slice.
Φ \Phi Φ is not propositional and S = ( Φ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) S = (\Phi,(\alpha_{1},\ldots,\alpha_{g}),(\beta_{1},\ldots,\beta_{g}),(S_{1},\ldots,S_{g}),\rho) S = ( Φ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) with g ≥ 1 g \ge1 g ≥ 1 , α j ∈ ⟨ V 1 ⟩ \alpha_{j} \in\langle V_{1}\rangle α j ∈ ⟨ V 1 ⟩ , β j ∈ ⟨ B ⟩ \beta_{j} \in\langle B\rangle β j ∈ ⟨ B ⟩ , each S j S_{j} S j a slice for Φ > 1 ∣ β j \Phi_{>1\mid\beta_{j}} Φ > 1 ∣ β j , and ρ \rho ρ a Res refutation or a placeholder, such that
(a) if Q 1 = ∃ Q_{1} = \exists Q 1 = ∃ , then α j ∪ β j \alpha_{j} \cup\beta_{j} α j ∪ β j satisfies Φ ∃ V 1 ∧ ⋀ a = 1 j − 1 G β a ∃ \Phi_{\exists V_{1}} \land\bigwedge_{a=1}^{j-1}G_{\beta_{a}}^{\exists} Φ ∃ V 1 ∧ ⋀ a = 1 j − 1 G β a ∃ for j = 1 , … , g j = 1,\ldots,g j = 1 , … , g . S S S is a refutation slice if S 1 , … , S g S_{1},\ldots,S_{g} S 1 , … , S g are refutation slices and ρ \rho ρ is a Res refutation of Φ ∃ V 1 ∧ ⋀ a = 1 g G β a ∃ \Phi_{\exists V_{1}} \land\bigwedge_{a=1}^{g}G_{\beta_{a}}^{\exists} Φ ∃ V 1 ∧ ⋀ a = 1 g G β a ∃ ; S S S is a verification slice if S 1 , … , S g − 1 S_{1},\ldots,S_{g-1} S 1 , … , S g − 1 are refutation slices and S g S_{g} S g is a verification slice (in this case ρ \rho ρ is a placeholder);
(b) if Q 1 = ∀ Q_{1} = \forall Q 1 = ∀ , then β j \beta_{j} β j is canonical for α j \alpha_{j} α j 4 and α j ∪ β j \alpha_{j} \cup\beta_{j} α j ∪ β j satisfies Φ ∀ V 1 ∧ ⋀ a = 1 j − 1 G β a ∀ \Phi_{\forall V_{1}} \land\bigwedge_{a=1}^{j-1}G_{\beta_{a}}^{\forall} Φ ∀ V 1 ∧ ⋀ a = 1 j − 1 G β a ∀ for j = 1 , … , g j = 1,\ldots,g j = 1 , … , g . S S S is a verification slice if S 1 , … , S g S_{1},\ldots,S_{g} S 1 , … , S g are verification slices and ρ \rho ρ is a Res refutation of Φ ∀ V 1 ∧ ⋀ a = 1 g G β a ∀ \Phi_{\forall V_{1}} \land\bigwedge_{a=1}^{g}G_{\beta_{a}}^{\forall} Φ ∀ V 1 ∧ ⋀ a = 1 g G β a ∀ ; S S S is a refutation slice if S 1 , … , S g − 1 S_{1},\ldots,S_{g-1} S 1 , … , S g − 1 are verification slices and S g S_{g} S g is a refutation slice (in this case ρ \rho ρ is a placeholder).
We set ∣ S ∣ : = g \lvert S\rvert:= g ∣ S ∣ := g and say that S S S refers to the slices S 1 , … , S g S_{1},\ldots,S_{g} S 1 , … , S g . Furthermore, we set Block ( S ) : = Q 1 V 1 \operatorname{Block}(S) := Q_{1}V_{1} Block ( S ) := Q 1 V 1 .
For the propositional slices of the first two cases we also set Block ( S ) : = ∃ V 1 \operatorname{Block}(S) := \exists V_{1} Block ( S ) := ∃ V 1 if the prefix of Φ \Phi Φ consists of the single block ∃ V 1 \exists V_{1} ∃ V 1 .
Note that we do not include computations of the SAT solver (i.e. ρ \rho ρ ) in the size of our slices. This is because first we want to concentrate on genuine QBF-hardness where SAT calls come for free [12 ] and second we do not specify the SAT solver that is used.
Next we define C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE proofs and distinguish between refutations (for false formulas) and verifications (for true formulas). Intuitively, a C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE proof for a QCNF Φ \Phi Φ is a sequence of slices, starting with the trivial cases of propositional formulas (one can consider these slices as axioms) and ending with a slice for Φ \Phi Φ itself. From an algorithmic perspective, a proof is generated from back to front. The last slices of a proof correspond to the first quantifier blocks, while the first slices belong to the last quantifier block.
Definition 10 (CAQEC O R E ^{\mathrm{CORE}} CORE proofs). A C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE proof κ \kappa κ of a Q C N F QCNF QC N F Φ \Phi Φ is a sequence of slices κ = ( S 1 , … , S e ) \kappa= (S_{1},\ldots,S_{e}) κ = ( S 1 , … , S e ) such that S e S_{e} S e is a slice for Φ \Phi Φ and each slice refers only to slices that occur earlier in the sequence. We set ∣ κ ∣ : = ∑ i = 1 e ∣ S i ∣ \lvert\kappa\rvert:= \sum_{i=1}^{e}\lvert S_{i}\rvert ∣ κ ∣ := ∑ i = 1 e ∣ S i ∣ . The proof κ \kappa κ is called a CAQEC O R E ^{\mathrm{CORE}} CORE refutation of Φ \Phi Φ if S e S_{e} S e is a refutation slice, and a CAQEC O R E ^{\mathrm{CORE}} CORE verification of Φ \Phi Φ if S e S_{e} S e is a verification slice.
We can check in polynomial time whether a given sequence of slices is a CAQEC o r e ^{\mathrm{Core}} Core refutation (or verification), since the conditions in Definition 9 are purely syntactic and whether a slice is a refutation or a verification slice only depends on the slices it refers to. Hence, CAQEC o r e ^{\mathrm{Core}} Core is a proof system in the sense of [22 ] as soon as we have shown soundness (Corollary 20) and completeness (Corollary 31).
It is straightforward to check that CAQEC o r e ^{\mathrm{Core}} Core proofs tightly model the basic algorithmic CAQE paradigm without failed assumptions.
Theorem 11. We call a CAQEC o r e ^{\mathrm{Core}} Core proof tree-like if every slice is referred to by at most one other slice. From each CAQE run without failed assumptions (that means the first solve procedure from [35 ] , where the SAT solver returns the maximal assignment b max b_{\max} b m a x of the auxiliary variables on universal blocks) we can efficiently extract a tree-like CAQEC o r e ^{\mathrm{Core}} Core proof. Conversely, each tree-like CAQEC o r e ^{\mathrm{Core}} Core proof can be efficiently transformed into such a CAQE run.
Proof Sketch. Consider the recursion tree of a run of the solve procedure on a QCNF Φ \Phi Φ . Each call of solve gets a formula Ψ \Psi Ψ . This is Φ \Phi Φ restricted by the assignments of the auxiliary variables that were chosen by the calls above it. In other words, Ψ = Φ > 1 ∣ β \Psi= \Phi_{>1}\vert_{\beta} Ψ = Φ > 1 ∣ β for the current β \beta β of the parent call (resp. Ψ = Φ \Psi= \Phi Ψ = Φ for the top-level call). If Ψ \Psi Ψ is propositional, the SAT solver returns either a Res refutation π \pi π or a satisfying assignment σ \sigma σ . This gives us the slice ( Ψ , π ) (\Psi,\pi) ( Ψ , π ) or ( Ψ , σ ) (\Psi,\sigma) ( Ψ , σ ) . In the other case, the call iterates. In the j t h j^{\mathrm{th}} j th iteration the SAT solver returns an assignment α j ∪ β j \alpha_j \cup\beta_j α j ∪ β j of the clausal abstraction Ψ Q 1 V 1 \Psi_{Q_1V_1} Ψ Q 1 V 1 together with the refinement clauses added so far, namely of Ψ Q 1 V 1 ∧ ⋀ a < j G β a ∃ \Psi_{Q_1V_1} \land\bigwedge_{a<j}G_{\beta_a}^{\exists} Ψ Q 1 V 1 ∧ ⋀ a < j G β a ∃ (resp. G β a ∀ G_{\beta_a}^{\forall} G β a ∀ ). After that, a recursive call on Ψ > 1 ∣ β j \Psi_{>1}\vert_{\beta_j} Ψ > 1 ∣ β j follows. The iteration stops for one of two reasons. Either the recursive call returns the “winning” result for the player of the first block (true in the existential case, false in the universal case), or the SAT solver reports that the abstraction with all refinement clauses is unsatisfiable. In the latter case the SAT solver returns a Res refutation ρ \rho ρ . If we collect Ψ \Psi Ψ , the α j , β j \alpha_j,\beta_j α j , β j , the slices of the recursive calls and ρ \rho ρ , we get a slice for Ψ \Psi Ψ . It follows directly from the description of the algorithm that this slice is a refutation slice if the call returns false and a verification slice if the call returns true. Now we list the slices of the recursion tree in post-order and obtain a CAQEC o r e ^{\mathrm{Core}} Core proof of Φ \Phi Φ , which is tree-like because every call has exactly one parent. Its size is the number of SAT calls that returned an assignment plus the number of calls on propositional formulas that returned a refutation (these correspond to the slices of the first kind in Definition 9, which have size 1). Note that the SAT solver can return different satisfying assignments of the same abstraction, hence the run is not unique. In the universal case we assume that the SAT solver returns b max b_{\max} b m a x , which is the canonical assignment.
For the other direction, let κ \kappa κ be a tree-like CAQEC o r e ^{\mathrm{Core}} Core proof of Φ \Phi Φ . We run solve with a SAT solver that answers each query with the assignment α j ∪ β j \alpha_j \cup\beta_j α j ∪ β j (resp. the refutation ρ \rho ρ , or the assignment σ \sigma σ ) that is stored in the corresponding slice of κ \kappa κ . By Definition 9, these answers are correct answers of the SAT solver, and the recursive calls by the run follow the references between the slices of κ \kappa κ . If κ \kappa κ is not tree-like, the same construction visits a slice once for each reference to it, so the run corresponds to the tree unfolding of κ \kappa κ . For a bounded number of quantifier blocks the unfolding is polynomial in ∣ κ ∣ \lvert\kappa\rvert ∣ κ ∣ , but in general it can be exponentially larger than κ \kappa κ .
CAQEC o r e ^{\mathrm{Core}} Core refutations only need to contain verification slices if the SAT solver did not succeed in immediately finding a β \beta β on a universal block such that Φ > 1 ∣ β \Phi_{>1}\vert_{\beta} Φ > 1 ∣ β is false. In most of our results we only need to consider a shortest CAQEC o r e ^{\mathrm{Core}} Core refutation, for which we can assume that slices for universal blocks only contain one assignment (namely the correct assignment).
This is because we get a smaller refutation if we delete the verification slices S 1 , … , S g − 1 S_1,\ldots,S_{g-1} S 1 , … , S g − 1 (and all slices only referred to by them) from a refutation slice of a universal block. That means for such results we can assume that C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutations do not contain verification slices.
Let us also point out a difference between the basic procedure and the implementation of CAQE. The implementation keeps one SAT solver per quantifier block for the whole run [35 Section IV.A.] . Hence, the refinement clauses that were learned in one call are also available in all later calls for the same block, even if these calls work on different restricted formulas. In Definition 9, however, the refutation ρ \rho ρ of a slice may only use the refinement clauses of this slice, in accordance with the pseudocode of the first solve procedure of [35 ] . We leave the analysis of this reuse of refinements – like that of failed assumptions – to future work. In particular, our lower bounds refer to the basic procedure.
▶ Example 12. Let Ψ \Psi Ψ be the QCNF with the prefix ∃ d ∀ e ∃ x 1 , x 2 , f ∀ u 1 , u 2 , w ∃ y 1 , y 2 , z \exists d\forall e\exists x_1,x_2,f\forall u_1,u_2,w\exists y_1,y_2,z ∃ d ∀ e ∃ x 1 , x 2 , f ∀ u 1 , u 2 , w ∃ y 1 , y 2 , z and the matrix
C 1 = d ∨ e ∨ x 1 ∨ u 1 ∨ y 1 C 2 = d ∨ e ∨ x 1 ∨ x 2 ∨ u 2 ∨ y 2 C 3 = d ∨ e ∨ y ˉ 1 ∨ y ˉ 2 C 4 = d ∨ e ∨ x ˉ 1 ∨ u ˉ 1 ∨ y 1 C 5 = d ∨ e ∨ x ˉ 1 ∨ u ˉ 2 ∨ y 2 C 6 = d ∨ e ∨ x ˉ 2 ∨ u ˉ 2 ∨ y 2 C 7 = d ˉ ∨ e ˉ ∨ f ∨ w ∨ z C 8 = d ˉ ∨ e ˉ ∨ f ˉ ∨ w ˉ ∨ z C 9 = d ˉ ∨ e ˉ ∨ z ˉ \begin{aligned}
C_1 &= d \lor e \lor x_1 \lor u_1 \lor y_1 & C_2 &= d \lor e \lor x_1 \lor x_2 \lor u_2 \lor y_2 & C_3 &= d \lor e \lor\bar{y}_1 \lor\bar{y}_2 \\
C_4 &= d \lor e \lor\bar{x}_1 \lor\bar{u}_1 \lor y_1 & C_5 &= d \lor e \lor\bar{x}_1 \lor\bar{u}_2 \lor y_2 & C_6 &= d \lor e \lor\bar{x}_2 \lor\bar{u}_2 \lor y_2 \\
C_7 &= \bar{d} \lor\bar{e} \lor f \lor w \lor z & C_8 &= \bar{d} \lor\bar{e} \lor\bar{f} \lor\bar{w} \lor z & C_9 &= \bar{d} \lor\bar{e} \lor\bar{z}
\end{aligned} C 1 C 4 C 7 = d ∨ e ∨ x 1 ∨ u 1 ∨ y 1 = d ∨ e ∨ x ˉ 1 ∨ u ˉ 1 ∨ y 1 = d ˉ ∨ e ˉ ∨ f ∨ w ∨ z C 2 C 5 C 8 = d ∨ e ∨ x 1 ∨ x 2 ∨ u 2 ∨ y 2 = d ∨ e ∨ x ˉ 1 ∨ u ˉ 2 ∨ y 2 = d ˉ ∨ e ˉ ∨ f ˉ ∨ w ˉ ∨ z C 3 C 6 C 9 = d ∨ e ∨ y ˉ 1 ∨ y ˉ 2 = d ∨ e ∨ x ˉ 2 ∨ u ˉ 2 ∨ y 2 = d ˉ ∨ e ˉ ∨ z ˉ
If we set d ↦ 0 d \mapsto0 d ↦ 0 and e ↦ 0 e \mapsto0 e ↦ 0 we end up with the formula from Example 2, and if we set d ↦ 1 d \mapsto1 d ↦ 1 and e ↦ 1 e \mapsto1 e ↦ 1 we end up with the first equality formula E q 1 \mathrm{Eq}_1 Eq 1 (cf. Definition 25 ).
The sequence of slices κ = ( S 1 , … , S 15 ) \kappa=(S_1,\ldots,S_{15}) κ = ( S 1 , … , S 15 ) is a C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutation of Ψ \Psi Ψ , with the slices defined as follows:
S 1 = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x ˉ 1 ∧ x ˉ 2 ∧ u ˉ 1 ∧ u ˉ 2 , π 1 ) S 2 = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x ˉ 1 ∧ x 2 ∧ u ˉ 1 ∧ u 2 , π 2 ) S 3 = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x 1 ∧ x ˉ 2 ∧ u 1 ∧ u 2 , π 3 ) S 4 = ( Ψ ∣ d ∧ e ∧ f ∧ w , π 4 ) S 5 = ( Ψ ∣ d ∧ e ∧ f ˉ ∧ w ˉ , π 5 ) S 6 = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x ˉ 1 ∧ x ˉ 2 , ( α 1 = u ˉ 1 ∧ u ˉ 2 ) , ( β 1 ) , ( S 1 ) , ρ 6 ) S 7 = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x ˉ 1 ∧ x 2 , ( α 1 = u ˉ 1 ∧ u 2 ) , ( β 1 ) , ( S 2 ) , ρ 7 ) S 8 = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x 1 ∧ x ˉ 2 , ( α 1 = u 1 ∧ u 2 ) , ( β 1 ) , ( S 3 ) , ρ 8 ) S 9 = ( Ψ ∣ d ∧ e ∧ f , ( α 1 = w ) , ( β 1 ) , ( S 4 ) , ρ 9 ) S 10 = ( Ψ ∣ d ∧ e ∧ f ˉ , ( α 1 = w ˉ ) , ( β 1 ) , ( S 5 ) , ρ 10 ) S 11 = ( Ψ ∣ d ˉ ∧ e ˉ , ( α 1 = x ˉ 1 ∧ x ˉ 2 , α 2 = x ˉ 1 ∧ x 2 , α 3 = x 1 ∧ x ˉ 2 ) , ( β 1 , β 2 , β 3 ) , ( S 6 , S 7 , S 8 ) , ρ 11 ) S 12 = ( Ψ ∣ d ∧ e , ( α 1 = [ f ] , α 2 = [ f ˉ ] ) , ( β 1 , β 2 ) , ( S 9 , S 10 ) , ρ 12 ) S 13 = ( Ψ ∣ d ˉ , ( α 1 = [ e ˉ ] ) , ( β 1 ) , ( S 11 ) , ρ 13 ) S 14 = ( Ψ ∣ d , ( α 1 = [ e ] ) , ( β 1 ) , ( S 12 ) , ρ 14 ) S 15 = ( Ψ , ( α 1 = [ d ˉ ] , α 2 = [ d ] ) , ( β 1 , β 2 ) , ( S 13 , S 14 ) , ρ 15 ) \begin{aligned}
S_1 &= (\Psi|_{\bar{d}\land\bar{e}\land\bar{x}_1\land\bar{x}_2\land\bar{u}_1\land\bar{u}_2},\pi_1) \\
S_2 &= (\Psi|_{\bar{d}\land\bar{e}\land\bar{x}_1\land x_2\land\bar{u}_1\land u_2},\pi_2) \\
S_3 &= (\Psi|_{\bar{d}\land\bar{e}\land x_1\land\bar{x}_2\land u_1\land u_2},\pi_3) \\
S_4 &= (\Psi|_{d\land e\land f\land w},\pi_4) \\
S_5 &= (\Psi|_{d\land e\land\bar{f}\land\bar{w}},\pi_5) \\
S_6 &= (\Psi|_{\bar{d}\land\bar{e}\land\bar{x}_1\land\bar{x}_2},(\alpha_1=\bar{u}_1\land\bar{u}_2),(\beta_1),(S_1),\rho_6) \\
S_7 &= (\Psi|_{\bar{d}\land\bar{e}\land\bar{x}_1\land x_2},(\alpha_1=\bar{u}_1\land u_2),(\beta_1),(S_2),\rho_7) \\
S_8 &= (\Psi|_{\bar{d}\land\bar{e}\land x_1\land\bar{x}_2},(\alpha_1=u_1\land u_2),(\beta_1),(S_3),\rho_8) \\
S_9 &= (\Psi|_{d\land e\land f},(\alpha_1=w),(\beta_1),(S_4),\rho_9) \\
S_{10} &= (\Psi|_{d\land e\land\bar{f}},(\alpha_1=\bar{w}),(\beta_1),(S_5),\rho_{10}) \\
S_{11} &= (\Psi|_{\bar{d}\land\bar{e}},(\alpha_1=\bar{x}_1\land\bar{x}_2,\alpha_2=\bar{x}_1\land x_2,\alpha_3=x_1\land\bar{x}_2),(\beta_1,\beta_2,\beta_3),(S_6,S_7,S_8),\rho_{11}) \\
S_{12} &= (\Psi|_{d\land e},(\alpha_1=[f],\alpha_2=[\bar{f}]),(\beta_1,\beta_2),(S_9,S_{10}),\rho_{12}) \\
S_{13} &= (\Psi|_{\bar{d}},(\alpha_1=[\bar{e}]),(\beta_1),(S_{11}),\rho_{13}) \\
S_{14} &= (\Psi|_d,(\alpha_1=[e]),(\beta_1),(S_{12}),\rho_{14}) \\
S_{15} &= (\Psi,(\alpha_1=[\bar{d}],\alpha_2=[d]),(\beta_1,\beta_2),(S_{13},S_{14}),\rho_{15})
\end{aligned} S 1 S 2 S 3 S 4 S 5 S 6 S 7 S 8 S 9 S 10 S 11 S 12 S 13 S 14 S 15 = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x ˉ 1 ∧ x ˉ 2 ∧ u ˉ 1 ∧ u ˉ 2 , π 1 ) = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x ˉ 1 ∧ x 2 ∧ u ˉ 1 ∧ u 2 , π 2 ) = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x 1 ∧ x ˉ 2 ∧ u 1 ∧ u 2 , π 3 ) = ( Ψ ∣ d ∧ e ∧ f ∧ w , π 4 ) = ( Ψ ∣ d ∧ e ∧ f ˉ ∧ w ˉ , π 5 ) = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x ˉ 1 ∧ x ˉ 2 , ( α 1 = u ˉ 1 ∧ u ˉ 2 ) , ( β 1 ) , ( S 1 ) , ρ 6 ) = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x ˉ 1 ∧ x 2 , ( α 1 = u ˉ 1 ∧ u 2 ) , ( β 1 ) , ( S 2 ) , ρ 7 ) = ( Ψ ∣ d ˉ ∧ e ˉ ∧ x 1 ∧ x ˉ 2 , ( α 1 = u 1 ∧ u 2 ) , ( β 1 ) , ( S 3 ) , ρ 8 ) = ( Ψ ∣ d ∧ e ∧ f , ( α 1 = w ) , ( β 1 ) , ( S 4 ) , ρ 9 ) = ( Ψ ∣ d ∧ e ∧ f ˉ , ( α 1 = w ˉ ) , ( β 1 ) , ( S 5 ) , ρ 10 ) = ( Ψ ∣ d ˉ ∧ e ˉ , ( α 1 = x ˉ 1 ∧ x ˉ 2 , α 2 = x ˉ 1 ∧ x 2 , α 3 = x 1 ∧ x ˉ 2 ) , ( β 1 , β 2 , β 3 ) , ( S 6 , S 7 , S 8 ) , ρ 11 ) = ( Ψ ∣ d ∧ e , ( α 1 = [ f ] , α 2 = [ f ˉ ]) , ( β 1 , β 2 ) , ( S 9 , S 10 ) , ρ 12 ) = ( Ψ ∣ d ˉ , ( α 1 = [ e ˉ ]) , ( β 1 ) , ( S 11 ) , ρ 13 ) = ( Ψ ∣ d , ( α 1 = [ e ]) , ( β 1 ) , ( S 12 ) , ρ 14 ) = ( Ψ , ( α 1 = [ d ˉ ] , α 2 = [ d ]) , ( β 1 , β 2 ) , ( S 13 , S 14 ) , ρ 15 )
To keep the example readable, we omit variables that no longer occur in the matrix of the corresponding restricted formula. Quantifier blocks that become empty in this way are dropped as well. For example, in S 12 S_{12} S 12 we only consider assignments of f f f because x 1 x_1 x 1 and x 2 x_2 x 2 no longer occur after the assignment d ∧ e d\land e d ∧ e . Strictly speaking, the omitted variables are assigned arbitrarily by the α \alpha α ’s of their blocks – for example, the α \alpha α ’s of S 11 S_{11} S 11 also assign f f f , and the α \alpha α ’s of S 6 , S 7 , S 8 S_6,S_7,S_8 S 6 , S 7 , S 8 also assign w w w . This has no effect on the restricted formulas. All assignments β \beta β are chosen canonically, such that they precisely indicate the effect of the corresponding α \alpha α .
For example, β 1 \beta_1 β 1 from the slice S 15 S_{15} S 15 is the assignment b 1 ∧ b 2 ∧ b 3 ∧ b 4 ∧ b 5 ∧ b 6 ∧ b ˉ 7 ∧ b ˉ 8 ∧ b ˉ 9 b_1 \land b_2 \land b_3 \land b_4 \land b_5 \land b_6 \land\bar{b}_7 \land\bar{b}_8 \land\bar{b}_9 b 1 ∧ b 2 ∧ b 3 ∧ b 4 ∧ b 5 ∧ b 6 ∧ b ˉ 7 ∧ b ˉ 8 ∧ b ˉ 9 . The π i \pi_i π i from the slices S 1 , … , S 5 S_1,\ldots,S_5 S 1 , … , S 5 are Res refutations of the corresponding formulas and ρ 11 , ρ 12 , ρ 15 \rho_{11},\rho_{12},\rho_{15} ρ 11 , ρ 12 , ρ 15 are Res refutations according to Definition 9, whereas ρ 6 , … , ρ 10 , ρ 13 , ρ 14 \rho_6,\ldots,\rho_{10},\rho_{13},\rho_{14} ρ 6 , … , ρ 10 , ρ 13 , ρ 14 are placeholders. Note that the slices of formulas starting with a universal quantifier, which are S 6 , S 7 , S 8 , S 9 , S 10 , S 13 , S 14 S_6,S_7,S_8,S_9,S_{10},S_{13},S_{14} S 6 , S 7 , S 8 , S 9 , S 10 , S 13 , S 14 , only consist of one α \alpha α and one β \beta β assignment. This is because we have always immediately picked the correct α \alpha α that represents the universal winning strategy. In practice we would potentially make a few incorrect guesses before we pick a correct α \alpha α .
We can depict CAQEC O R E ^{\mathrm{CORE}} CORE proofs in a graph representation, where each node is a slice and we connect slices that refer to each other. Below we show a graph representation of κ \kappa κ . Each edge is labelled with the assignment α \alpha α that leads to the recursive call. Slices with recursive calls are drawn in blue, and the propositional slices, which are handled by the SAT solver, are drawn in yellow. The rows correspond to the quantifier blocks of Ψ \Psi Ψ .
Figure 3 (PDF p. 14).
One can observe that S 11 S_{11} S 11 only splits into three slices instead of four, although we have two variables remaining in the first block, namely x 1 x_1 x 1 and x 2 x_2 x 2 . This is because only three of these four assignments are optimal assignments. As described in Example 2, the assignment x 1 ∧ x 2 x_1 \land x_2 x 1 ∧ x 2 does not need to be considered since it is worse than the previous assignment x 1 ∧ x ˉ 2 x_1 \land\bar{x}_2 x 1 ∧ x ˉ 2 , which is already handled in S 8 S_8 S 8 . This ensures that our proof, and therefore the graph representation, does not necessarily blow up – although there are formulas where indeed all of the assignments need to be tested because all of them are optimal.
Note that graph representations of CAQEC O R E ^{\mathrm{CORE}} CORE proofs do not need to be trees. If we have two slices S S S and S ′ S' S ′ for the same formula, then we can omit S ′ S' S ′ and make all slices that refer to S ′ S' S ′ refer to S S S instead.
The following two lemmas relate the formulas Φ > 1 ∣ β h \Phi_{>1\mid\beta_h} Φ > 1 ∣ β h of the recursive calls to the restrictions Φ ∣ α h \Phi\mid_{\alpha_h} Φ ∣ α h of the original formula. We will use them in the whole paper.
Lemma 13. Let Φ \Phi Φ be a QCNF that starts with a universal quantifier ∀ V 1 \forall V_1 ∀ V 1 , and let α h ∈ ⟨ V 1 ⟩ \alpha_h \in\langle V_1\rangle α h ∈ ⟨ V 1 ⟩ and β h ∈ ⟨ B ⟩ \beta_h \in\langle B\rangle β h ∈ ⟨ B ⟩ be such that α h ∪ β h \alpha_h \cup\beta_h α h ∪ β h satisfies Φ ∀ V 1 \Phi_{\forall V_1} Φ ∀ V 1 (for example the assignments of a slice for Φ \Phi Φ , cf. Definition 9). Then we have Φ > 1 ∣ β h ⊆ Φ ∣ α h \Phi_{>1\mid\beta_h} \subseteq\Phi\mid_{\alpha_h} Φ > 1 ∣ β h ⊆ Φ ∣ α h . Furthermore, if β h \beta_h β h is canonical for α h \alpha_h α h (in which case α h ∪ β h \alpha_h \cup\beta_h α h ∪ β h satisfies Φ ∀ V 1 \Phi_{\forall V_1} Φ ∀ V 1 by Lemma 5), then we even have Φ > 1 ∣ β h = Φ ∣ α h \Phi_{>1\mid\beta_h} = \Phi\mid_{\alpha_h} Φ > 1 ∣ β h = Φ ∣ α h . In particular, the latter holds for all assignments of a slice for Φ \Phi Φ due to Definition 9.
Proof. Both formulas have the same prefix. Let C q , > 1 ∣ β h C_{q,>1\mid\beta_h} C q , > 1 ∣ β h be a clause of Φ > 1 ∣ β h \Phi_{>1\mid\beta_h} Φ > 1 ∣ β h , that is, β h ( b q ) = 1 \beta_h(b_q)=1 β h ( b q ) = 1 . Since α h ∪ β h \alpha_h\cup\beta_h α h ∪ β h satisfies C q , 1 = ⋀ ℓ ∈ C q , var ( ℓ ) ∈ V 1 ( ℓ ˉ ∨ b ˉ q ) C_{q,1}=\bigwedge_{\ell\in C_q,\ \operatorname{var}(\ell)\in V_1}(\bar{\ell}\lor\bar{b}_q) C q , 1 = ⋀ ℓ ∈ C q , var ( ℓ ) ∈ V 1 ( ℓ ˉ ∨ b ˉ q ) , the assignment α h \alpha_h α h falsifies all literals of C q C_q C q over V 1 V_1 V 1 . Hence, α h \alpha_h α h does not satisfy C q C_q C q , and
C q , > 1 ∣ β h = ⋁ ℓ ∈ C q var ( ℓ ) ∉ V 1 ℓ = C q ∣ α h C_{q,>1\mid\beta_h}=\bigvee_{\substack{\ell\in C_q\\\operatorname{var}(\ell)\notin V_1}}\ell=C_q\mid_{\alpha_h} C q , > 1 ∣ β h = ℓ ∈ C q var ( ℓ ) ∈ / V 1 ⋁ ℓ = C q ∣ α h
is a clause of Φ ∣ α h \Phi|_{\alpha_h} Φ ∣ α h . If β h \beta_h β h is canonical for α h \alpha_h α h , then we have β h ( b q ) = 1 \beta_h(b_q)=1 β h ( b q ) = 1 for every clause C q C_q C q that is not satisfied by α h \alpha_h α h . So conversely every clause C q ∣ α h C_q|_{\alpha_h} C q ∣ α h of Φ ∣ α h \Phi|_{\alpha_h} Φ ∣ α h is the clause C q , > 1 ∣ β h C_{q,>1}|_{\beta_h} C q , > 1 ∣ β h of Φ > 1 ∣ β h \Phi_{>1}|_{\beta_h} Φ > 1 ∣ β h .
Lemma 14. Let Φ \Phi Φ be a Q C N F QCNF QC N F that starts with an existential quantifier ∃ V 1 \exists V_1 ∃ V 1 , and let α h ∈ ⟨ V 1 ⟩ \alpha_h\in\langle V_1\rangle α h ∈ ⟨ V 1 ⟩ and β h ∈ ⟨ B ⟩ \beta_h\in\langle B\rangle β h ∈ ⟨ B ⟩ be such that α h ∪ β h \alpha_h\cup\beta_h α h ∪ β h satisfies Φ ∃ V 1 \Phi_{\exists V_1} Φ ∃ V 1 (for example the assignments of a slice for Φ \Phi Φ ). Then we have Φ > 1 ∣ β h ⊇ Φ ∣ α h \Phi_{>1}|_{\beta_h}\supseteq\Phi|_{\alpha_h} Φ > 1 ∣ β h ⊇ Φ ∣ α h . Furthermore, if β h \beta_h β h is canonical for α h \alpha_h α h , we even have Φ > 1 ∣ β h = Φ ∣ α h \Phi_{>1}|_{\beta_h}=\Phi|_{\alpha_h} Φ > 1 ∣ β h = Φ ∣ α h .
Proof. Both formulas have the same prefix. Let C q ∣ α h C_q|_{\alpha_h} C q ∣ α h be a clause of Φ ∣ α h \Phi|_{\alpha_h} Φ ∣ α h , that is, α h \alpha_h α h does not satisfy C q C_q C q . Then α h \alpha_h α h falsifies all literals of C q C_q C q over V 1 V_1 V 1 . Since α h ∪ β h \alpha_h\cup\beta_h α h ∪ β h satisfies C q , 1 = ⋁ ℓ ∈ C q , var ( ℓ ) ∈ V 1 ℓ ∨ b q C_{q,1}=\bigvee_{\ell\in C_q,\ \operatorname{var}(\ell)\in V_1}\ell\vee b_q C q , 1 = ⋁ ℓ ∈ C q , var ( ℓ ) ∈ V 1 ℓ ∨ b q , we get β h ( b q ) = 1 \beta_h(b_q)=1 β h ( b q ) = 1 . Hence,
C q ∣ α h = ⋁ ℓ ∈ C q var ( ℓ ) ∉ V 1 ℓ = C q , > 1 ∣ β h C_q|_{\alpha_h}=\bigvee_{\substack{\ell\in C_q\\ \operatorname{var}(\ell)\notin V_1}}\ell=C_{q,>1}|_{\beta_h} C q ∣ α h = ℓ ∈ C q var ( ℓ ) ∈ / V 1 ⋁ ℓ = C q , > 1 ∣ β h is a clause of Φ > 1 ∣ β h \Phi_{>1}|_{\beta_h} Φ > 1 ∣ β h . If β h \beta_h β h is canonical for α h \alpha_h α h , then β h ( b q ) = 1 \beta_h(b_q)=1 β h ( b q ) = 1 only holds for clauses C q C_q C q that are not satisfied by α h \alpha_h α h , so Φ > 1 ∣ β h \Phi_{>1}|_{\beta_h} Φ > 1 ∣ β h contains no further clauses.
Next we recall the ∀ \forall ∀ Red + Res proof system introduced in [38 ] to analyse CEGAR-type solving. We will show that ∀ \forall ∀ Red + Res, after adding an NP oracle, simulates CAQEC O R E ^{\mathrm{CORE}} CORE proofs, but is exponentially stronger (Theorem 34 ).
Definition 15 (∀ \forall ∀ Red + Res, [38 ] ). Let Φ = Q 1 V 1 … Q n V n ⋅ C 1 ∧ … ∧ C k \Phi=Q_1V_1\ldots Q_nV_n\cdot C_1\land\ldots\land C_k Φ = Q 1 V 1 … Q n V n ⋅ C 1 ∧ … ∧ C k with Q i ∈ { ∃ , ∀ } Q_i\in\{\exists,\forall\} Q i ∈ { ∃ , ∀ } for each i i i and Q j ≠ Q j + 1 Q_j\ne Q_{j+1} Q j = Q j + 1 for each j j j . A ∀ \forall ∀ Red + Res proof of Φ \Phi Φ is a sequence of sets P a \mathcal{P}^a P a , where P ⊆ { 1 , … , k } \mathcal{P}\subseteq\{1,\ldots,k\} P ⊆ { 1 , … , k } and a ∈ { 0 , 1 , … , n } a\in\{0,1,\ldots,n\} a ∈ { 0 , 1 , … , n } , derived by the following three rules: the resolution rule P 1 a … P g a ( P 1 ∪ … ∪ P g ) a − 1 π \frac{\mathcal{P}_1^a\ldots\mathcal{P}_g^a}{(\mathcal{P}_1\cup\ldots\cup\mathcal{P}_g)^{a-1}}\ \pi ( P 1 ∪ … ∪ P g ) a − 1 P 1 a … P g a π where Q a = ∃ Q_a=\exists Q a = ∃ and π \pi π is a Res refutation of ⋀ h = 1 g ⋁ q ∈ P h C q ∣ V a \bigwedge_{h=1}^{g}\bigvee_{q\in\mathcal{P}_h}C_q|_{V_a} ⋀ h = 1 g ⋁ q ∈ P h C q ∣ V a , the reduction rule P a P a − 1 \frac{\mathcal{P}^a}{\mathcal{P}^{a-1}} P a − 1 P a where Q a = ∀ Q_a=\forall Q a = ∀ and there is no u ∈ V a u\in V_a u ∈ V a with u ∈ ⋁ q ∈ P C q ∣ V a u\in\bigvee_{q\in\mathcal{P}}C_q|_{V_a} u ∈ ⋁ q ∈ P C q ∣ V a and u ˉ ∈ ⋁ q ∈ P C q ∣ V a \bar{u}\in\bigvee_{q\in\mathcal{P}}C_q|_{V_a} u ˉ ∈ ⋁ q ∈ P C q ∣ V a , and the axiom rule { q } n { q } n \frac{\{q\}^n}{\phantom{\{q\}^n}} { q } n { q } n with q ∈ { 1 , … , k } q\in\{1,\ldots,k\} q ∈ { 1 , … , k } .
A ∀ \forall ∀ Red + Res proof is called a ∀ \forall ∀ Red + Res refutation if we derived P 0 \mathcal{P}^0 P 0 for some P ⊆ { 1 , … , k } \mathcal{P}\subseteq\{1,\ldots,k\} P ⊆ { 1 , … , k } . The size of a ∀ \forall ∀ Red + Res proof is the number of sets P a \mathcal{P}^a P a and the sum of the sizes of all Res refutations π \pi π used in the proof. We call ∀ \forall ∀ Red + ResN P ^{\mathrm{NP}} NP the version of ∀ \forall ∀ Red + Res in which the π \pi π are not counted.
For the sake of convenience, we will often write P V a \mathcal{P}^{V_a} P V a instead of P a \mathcal{P}^a P a for a > 0 a>0 a > 0 . That means the superscript names the block over which the next rule is applied. In the following proof we consider a Q C N F QCNF QC N F with the prefix ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 \forall U_0\exists X_1\forall U_1\ldots\exists X_m\forall U_m\exists X_{m+1} ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 , where U 0 U_0 U 0 might be empty. Then the line P 0 \mathcal{P}^0 P 0 is P U 0 \mathcal{P}^{U_0} P U 0 if U 0 U_0 U 0 is empty, and if not, we write P X 0 \mathcal{P}^{X_0} P X 0 for it.
Theorem 16. Let κ \kappa κ be a CAQEC O R E ^{\mathrm{CORE}} CORE refutation of a Q C N F QCNF QC N F Φ \Phi Φ with k k k clauses. Then we can efficiently construct a ∀ \forall ∀ Red + ResN P ^{\mathrm{NP}} NP refutation ν \nu ν of Φ \Phi Φ with ∣ ν ∣ ≤ 2 ∣ κ ∣ + k |\nu|\leq2|\kappa|+k ∣ ν ∣ ≤ 2∣ κ ∣ + k . Hence, ∀ \forall ∀ Red + ResN P ^{\mathrm{NP}} NP p p p -simulates CAQEC O R E ^{\mathrm{CORE}} CORE on false Q C N F s QCNFs QC N F s .
Proof. Let Φ \Phi Φ have the prefix ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 \forall U_0\exists X_1\forall U_1\ldots\exists X_m\forall U_m\exists X_{m+1} ∀ U 0 ∃ X 1 ∀ U 1 … ∃ X m ∀ U m ∃ X m + 1 , where U 0 U_0 U 0 is allowed to be empty, and the clauses C 1 , … , C k C_1,\ldots,C_k C 1 , … , C k . Let κ = ( S 1 , … , S e ) \kappa=(S_1,\ldots,S_e) κ = ( S 1 , … , S e ) and w.l.o.g. let κ \kappa κ only consist of relevant slices (meaning that all slices have to be connected to S e S_e S e somehow) and let each slice of a universal block contain only one assignment, hence referring to one further slice only (cf. the discussion after Theorem 11). In particular, all slices in κ \kappa κ are refutation slices. The formulas of the axiom slices S = ( Ψ , π ) S=(\Psi,\pi) S = ( Ψ , π ) have the prefix ∃ X m + 1 \exists X_{m+1} ∃ X m + 1 , and we set Block(S S S ) : = ∃ X m + 1 := \exists X_{m+1} := ∃ X m + 1 for them. Now let S ∈ κ S \in\kappa S ∈ κ be a slice with Block(S S S ) = ∃ X i = \exists X_i = ∃ X i for some i ∈ { 1 , … , m + 1 } i \in\{1,\ldots,m+1\} i ∈ { 1 , … , m + 1 } . We will define a set P S ⊆ { 1 , … , k } P_S \subseteq\{1,\ldots,k\} P S ⊆ { 1 , … , k } of clause indices by induction on m + 1 − i m+1-i m + 1 − i and show that the line P S U i − 1 P_S^{U_{i-1}} P S U i − 1 can be derived in ∀ \forall ∀ Red + + + Res. During the construction we maintain the following invariant:
( ∗ ) (*) ( ∗ ) If S S S is a slice for the formula Ψ \Psi Ψ , then every q ∈ P S q \in P_S q ∈ P S is the index of a clause that is present in Ψ \Psi Ψ .
Note that if C q C_q C q is present in Ψ \Psi Ψ , then the clause of Ψ \Psi Ψ with index q q q is simply C q C_q C q without the literals of the variables that are already assigned. In particular, its literals from X i X_i X i are precisely the literals of C q ∣ X i C_q|_{X_i} C q ∣ X i .
The invariant allows us to continue with a reduction step. If S S S is not the last slice of κ \kappa κ , then it is referred to by a slice ( Ψ ′ , ( α ′ ) , ( β ′ ) , ( S ) , ρ ′ ) (\Psi',(\alpha'),(\beta'),(S),\rho') ( Ψ ′ , ( α ′ ) , ( β ′ ) , ( S ) , ρ ′ ) of the block ∀ U i − 1 \forall U_{i-1} ∀ U i − 1 . As β ′ \beta' β ′ is canonical for α ′ \alpha' α ′ , we have Ψ = Ψ ′ ∣ β ′ = Ψ ′ ∣ α ′ \Psi=\Psi'|_{\beta'}=\Psi'|_{\alpha'} Ψ = Ψ ′ ∣ β ′ = Ψ ′ ∣ α ′ by Lemma 13. By ( ∗ ) (*) ( ∗ ) , for every q ∈ P S q\in P_S q ∈ P S the clause of Ψ ′ \Psi' Ψ ′ with index q q q is not satisfied by α ′ \alpha' α ′ , which means that α ′ \alpha' α ′ falsifies all literals of C q C_q C q from U i − 1 U_{i-1} U i − 1 . Hence, the clauses C q C_q C q with q ∈ P S q\in P_S q ∈ P S do not contain complementary literals from U i − 1 U_{i-1} U i − 1 , and the reduction step P S U i − 1 P S X i − 1 \frac{P_S^{U_{i-1}}}{P_S^{X_{i-1}}} P S X i − 1 P S U i − 1 is valid.
For i = m + 1 i=m+1 i = m + 1 , the slice S S S is an axiom slice of the form S = ( Ψ , π ) S=(\Psi,\pi) S = ( Ψ , π ) , where π \pi π is a Res refutation of the propositional formula Ψ \Psi Ψ . Let P S : = { q 1 , … , q a } P_S:=\{q_1,\ldots,q_a\} P S := { q 1 , … , q a } be the set of indices of all clauses that are present in Ψ \Psi Ψ . As we have seen above, Ψ = C q 1 ∣ X m + 1 ∧ ⋯ ∧ C q a ∣ X m + 1 \Psi=C_{q_1}|_{X_{m+1}}\land\cdots\land C_{q_a}|_{X_{m+1}} Ψ = C q 1 ∣ X m + 1 ∧ ⋯ ∧ C q a ∣ X m + 1 . Therefore,
{ q 1 } X m + 1 ⋯ { q a } X m + 1 π { q 1 , … , q a } U m \frac{\{q_1\}^{X_{m+1}}\ \cdots\ \{q_a\}^{X_{m+1}}\quad\pi}{\{q_1,\ldots,q_a\}^{U_m}} { q 1 , … , q a } U m { q 1 } X m + 1 ⋯ { q a } X m + 1 π
is a valid resolution step in ∀ \forall ∀ Red + + + Res, where the lines { q h } X m + 1 \{q_h\}^{X_{m+1}} { q h } X m + 1 are axioms. For i ≤ m i\leq m i ≤ m , let S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S j 1 , … , S j g ) , ρ ) S=(\Psi,(\alpha_1,\ldots,\alpha_g),(\beta_1,\ldots,\beta_g),(S_{j_1},\ldots,S_{j_g}),\rho) S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S j 1 , … , S j g ) , ρ ) with Block(S S S ) = ∃ X i =\exists X_i = ∃ X i . This slice refers to the slices S j 1 , … , S j g S_{j_1},\ldots,S_{j_g} S j 1 , … , S j g with Block(S j 1 ) = ⋯ = S_{j_1})=\cdots= S j 1 ) = ⋯ = Block($S_{j_g}) = ∀ U i =\forall U_i = ∀ U i . Each S j h S_{j_h} S j h refers to exactly one slice S j h ′ S'_{j_h} S j h ′ with Block(S j h ′ ) = ∃ X i + 1 S'_{j_h})=\exists X_{i+1} S j h ′ ) = ∃ X i + 1 . Inductively, we have already constructed the sets P S j 1 ′ , … , P S j g ′ P_{S'_{j_1}},\ldots,P_{S'_{j_g}} P S j 1 ′ , … , P S j g ′ and derived the lines P S j h ′ U i P_{S'_{j_h}}^{U_i} P S j h ′ U i , and with the reduction steps from above also the lines P S j h ′ X i P_{S'_{j_h}}^{X_i} P S j h ′ X i . We set P S : = ⋃ h = 1 g P S j h ′ P_S:=\bigcup_{h=1}^gP_{S'_{j_h}} P S := ⋃ h = 1 g P S j h ′ . The invariant ( ∗ ) (*) ( ∗ ) holds for P S P_S P S because each clause that is present in the formula of S j h ′ S'_{j_h} S j h ′ is also present in Ψ \Psi Ψ .
We claim that ⋀ h = 1 g ⋁ q ∈ P S j h ′ C q ∣ X i \bigwedge_{h=1}^g\bigvee_{q\in P_{S'_{j_h}}}C_q|_{X_i} ⋀ h = 1 g ⋁ q ∈ P S j h ′ C q ∣ X i is unsatisfiable. Assume for the sake of contradiction that this formula is satisfied by an assignment α ∈ ⟨ X i ⟩ \alpha\in\langle X_i\rangle α ∈ ⟨ X i ⟩ . Let β ∈ ⟨ B ⟩ \beta\in\langle B\rangle β ∈ ⟨ B ⟩ be the assignment that is canonical for α \alpha α with respect to Ψ \Psi Ψ . Due to Lemma 5, α ∪ β \alpha\cup\beta α ∪ β satisfies Ψ ∃ X i \Psi_{\exists X_i} Ψ ∃ X i . Now, for each h h h there exists some q ∈ P S j h ′ q\in P_{S'_{j_h}} q ∈ P S j h ′ such that α \alpha α satisfies C q ∣ X i C_q|_{X_i} C q ∣ X i , and hence the clause C q C_q C q of Ψ \Psi Ψ . Therefore β ( b q ) = 0 \beta(b_q)=0 β ( b q ) = 0 . On the other hand, C q C_q C q is present in the formula of S j h ′ S'_{j_h} S j h ′ , and this formula is a subformula of Ψ ∣ β h \Psi_{\rvert\beta_h} Ψ ∣ β h , so β h ( b q ) = 1 \beta_h(b_q)=1 β h ( b q ) = 1 . Using Lemma 8, we conclude that β \beta β satisfies G β h 3 G_{\beta_h}^3 G β h 3 for all h = 1 , … , g h=1,\ldots,g h = 1 , … , g . But this means that α ∪ β \alpha\cup\beta α ∪ β satisfies Ψ ∃ X i ∧ ⋀ h = 1 g G β h 3 \Psi_{\exists X_i}\land\bigwedge_{h=1}^gG_{\beta_h}^3 Ψ ∃ X i ∧ ⋀ h = 1 g G β h 3 , which contradicts the existence of the Res refutation ρ \rho ρ in the refutation slice S S S .
Hence, ⋀ h = 1 g ⋁ q ∈ P S j h ′ C q ∣ X i \bigwedge_{h=1}^g\bigvee_{q\in P_{S'_{j_h}}}C_q|_{X_i} ⋀ h = 1 g ⋁ q ∈ P S j h ′ C q ∣ X i is unsatisfiable and it has a Res refutation π \pi π .5 In particular,
P S j 1 ′ X i ⋯ P S j g ′ X i π ( ⋃ h = 1 g P S j h ′ ) U i − 1 \frac{P_{S'_{j_1}}^{X_i}\ \cdots\ P_{S'_{j_g}}^{X_i}\quad\pi}{\left(\bigcup_{h=1}^gP_{S'_{j_h}}\right)^{U_{i-1}}} ( ⋃ h = 1 g P S j h ′ ) U i − 1 P S j 1 ′ X i ⋯ P S j g ′ X i π
is a valid resolution step in ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res .
Finally, let S S S be the slice of the block ∃ X 1 \exists X_1 ∃ X 1 in κ \kappa κ . If U 0 U_0 U 0 is empty, then S S S is the last slice of κ \kappa κ , and the line P S U 0 = P S 0 P_S^{U_0}=P_S^0 P S U 0 = P S 0 completes the ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res refutation of Φ \Phi Φ . If not, then S S S is referred to by the last slice of κ \kappa κ , and the reduction step from above gives us P S X 0 = P S 0 P_S^{X_0}=P_S^0 P S X 0 = P S 0 .
It remains to consider the size. The refutation ν \nu ν uses at most k k k axioms and at most two further lines for each slice of an existential block in κ \kappa κ . Since each slice has size at least 1, we get ∣ ν ∣ ≤ 2 ∣ κ ∣ + k |\nu|\leq2|\kappa|+k ∣ ν ∣ ≤ 2∣ κ ∣ + k . Obviously, the construction can be computed in polynomial time – note that the refutations π \pi π are not part of a ∀ R e d + R e s N P \forall\mathrm{Red}+\mathrm{Res}^{\mathrm{NP}} ∀ Red + Res NP refutation.
We would like to comment that it is unclear whether ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res simulates the non-oracle version of C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core , or even the general algorithmic CAQE paradigm. While the Res refutations π \pi π in ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res proofs somehow correspond to the refutations ρ \rho ρ generated by a SAT solver during a C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core run, there might not be a polynomial-time transformation between those two. The reason is that the latter uses the auxiliary B B B -variables from C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core , while the former does not. Resolving over those B B B -variables first would effectively eliminate them, but this might result in an exponential blow-up in the proof size. On the other hand, redefining ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res by allowing B B B -variables in the refutations π \pi π could cause other problems: In [38 ] it was proven that ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res is equivalent to level-ordered Q-Res. For the first direction, one can argue that all π \pi π simply have to be stuck together in order to obtain a level-ordered Q-Res refutation. Allowing B B B -variables in π \pi π could render this simulation impossible.
We avoid these potential issues by adding an NP oracle and disregarding Res refutations both for ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res and for C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core . Whether the simulation of C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core by ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res also holds without NP oracles remains open. Note that such a simulation would not directly imply Theorem 16, since a C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutation of small oracle size may contain large Res refutations ρ \rho ρ .
Strategy extraction and soundness of CAQE
In this section we will show how to perform strategy extraction on C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proofs. Strategy extraction is a common technique not only for obtaining lower bounds, but also for proving the soundness of a proof system. More precisely, from a given C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutation we extract a universal winning strategy, whereas from a C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core verification we extract an existential winning strategy. This results in strategy size as a lower bound for the size of C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proofs.
From now on, if not otherwise specified, we will always assume that Φ \Phi Φ is a QCNF with prefix ∀ U 0 ∃ X 1 ∀ U 1 ∃ X 2 ∀ U 2 … ∃ X m ∀ U m ∃ X m + 1 \forall U_0\exists X_1\forall U_1\exists X_2\forall U_2\ldots\exists X_m\forall U_m\exists X_{m+1} ∀ U 0 ∃ X 1 ∀ U 1 ∃ X 2 ∀ U 2 … ∃ X m ∀ U m ∃ X m + 1 , where U 0 U_0 U 0 is allowed to be empty. For the sake of readability, we will denote X : = X 1 ∪ … ∪ X m + 1 X:=X_1\cup\ldots\cup X_{m+1} X := X 1 ∪ … ∪ X m + 1 and U : = U 0 ∪ … ∪ U m U:=U_0\cup\ldots\cup U_m U := U 0 ∪ … ∪ U m . Furthermore, for a slice S S S of a C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proof of Φ \Phi Φ with Block ( S ) = Q V \operatorname{Block}(S)=QV Block ( S ) = Q V we simply write Block ( S ) = V \operatorname{Block}(S)=V Block ( S ) = V . In particular, Block ( S ) = X m + 1 \operatorname{Block}(S)=X_{m+1} Block ( S ) = X m + 1 for the propositional slices S = ( Ψ , π ) S=(\Psi,\pi) S = ( Ψ , π ) and S = ( Ψ , σ ) S=(\Psi,\sigma) S = ( Ψ , σ ) , whose formula Ψ \Psi Ψ has the prefix ∃ X m + 1 \exists X_{m+1} ∃ X m + 1 (cf. Definition 9).
Both extraction procedures rest on the following lemma. It says that the assignments of a refutation slice of an existential block cover all assignments of this block, and dually for verification slices of universal blocks.
Lemma 17. Let S S S be a slice of the form ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) (\Psi,(\alpha_1,\ldots,\alpha_g),(\beta_1,\ldots,\beta_g),(S_1,\ldots,S_g),\rho) ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) with Block ( S ) = Q V \operatorname{Block}(S)=QV Block ( S ) = Q V . Furthermore, let γ ∈ ⟨ V ⟩ \gamma\in\langle V\rangle γ ∈ ⟨ V ⟩ and let β γ \beta_\gamma β γ be canonical for γ \gamma γ with respect to Ψ \Psi Ψ .
If Q = ∃ Q=\exists Q = ∃ and S S S is a refutation slice, then there is some h h h with β h − 1 ( 1 ) ⊆ β γ − 1 ( 1 ) \beta_h^{-1}(1)\subseteq\beta_\gamma^{-1}(1) β h − 1 ( 1 ) ⊆ β γ − 1 ( 1 ) . For every such h h h we have Ψ > 1 ∣ β h ⊆ Ψ ∣ γ \Psi_{>1|\beta_h}\subseteq\Psi|_\gamma Ψ > 1∣ β h ⊆ Ψ ∣ γ , and every clause of Ψ \Psi Ψ that is satisfied by γ \gamma γ is also satisfied by α h \alpha_h α h .
If Q = ∀ Q = \forall Q = ∀ and S S S is a verification slice, then there is some h h h with β h − 1 ( 0 ) ⊆ β γ − 1 ( 0 ) \beta_h^{-1}(0) \subseteq\beta_\gamma^{-1}(0) β h − 1 ( 0 ) ⊆ β γ − 1 ( 0 ) . For every such h h h we have Ψ > 1 ∣ β h ⊇ Ψ ∣ γ \Psi_{>1|\beta_h} \supseteq\Psi|_\gamma Ψ > 1∣ β h ⊇ Ψ ∣ γ , and every clause of Ψ \Psi Ψ that is satisfied by α h \alpha_h α h is also satisfied by γ \gamma γ . In both cases, if α h \alpha_h α h and γ \gamma γ satisfy the same clauses of Ψ \Psi Ψ , then β h = β γ \beta_h = \beta_\gamma β h = β γ and therefore Ψ > 1 ∣ β h = Ψ ∣ γ \Psi_{>1|\beta_h} = \Psi|_\gamma Ψ > 1∣ β h = Ψ ∣ γ .
Proof. Let Q = ∃ Q = \exists Q = ∃ . If there were no such h h h , then β γ \beta_\gamma β γ would satisfy ⋀ a = 1 g G β a ∃ \bigwedge_{a=1}^{g} G_{\beta_a}^{\exists} ⋀ a = 1 g G β a ∃ by Lemma 8 . As γ ∪ β γ \gamma\cup\beta_\gamma γ ∪ β γ satisfies Ψ ∃ V \Psi_{\exists V} Ψ ∃ V by Lemma 5 , this contradicts the existence of the refutation ρ \rho ρ of Ψ ∃ V ∧ ⋀ a = 1 g G β a ∃ \Psi_{\exists V} \land\bigwedge_{a=1}^{g} G_{\beta_a}^{\exists} Ψ ∃ V ∧ ⋀ a = 1 g G β a ∃ . Now let h h h be as in the statement. Then every clause of Ψ > 1 ∣ β h \Psi_{>1|\beta_h} Ψ > 1∣ β h is a clause of Ψ > 1 ∣ β γ \Psi_{>1|\beta_\gamma} Ψ > 1∣ β γ , and we have Ψ > 1 ∣ β γ = Ψ ∣ γ \Psi_{>1|\beta_\gamma} = \Psi|_\gamma Ψ > 1∣ β γ = Ψ ∣ γ by Lemma 14 . If γ \gamma γ satisfies C r C_r C r , then β γ ( b r ) = 0 \beta_\gamma(b_r) = 0 β γ ( b r ) = 0 and therefore β h ( b r ) = 0 \beta_h(b_r) = 0 β h ( b r ) = 0 . Since α h ∪ β h \alpha_h \cup\beta_h α h ∪ β h satisfies C r , 1 C_{r,1} C r , 1 , the assignment α h \alpha_h α h satisfies C r C_r C r . If α h \alpha_h α h and γ \gamma γ satisfy the same clauses, then conversely β h ( b r ) = 0 \beta_h(b_r) = 0 β h ( b r ) = 0 implies that α h \alpha_h α h and hence γ \gamma γ satisfies C r C_r C r , which means β γ ( b r ) = 0 \beta_\gamma(b_r) = 0 β γ ( b r ) = 0 . So we get β h = β γ \beta_h = \beta_\gamma β h = β γ .
The case Q = ∀ Q = \forall Q = ∀ is dual. Here the existence of h h h follows in the same way with the clauses G β a ∀ G_{\beta_a}^{\forall} G β a ∀ , and the assignments β h \beta_h β h are canonical by Definition 9 . Hence, β h − 1 ( 0 ) ⊆ β γ − 1 ( 0 ) \beta_h^{-1}(0) \subseteq\beta_\gamma^{-1}(0) β h − 1 ( 0 ) ⊆ β γ − 1 ( 0 ) says that every clause that is satisfied by α h \alpha_h α h is also satisfied by γ \gamma γ . Moreover, every clause of Ψ > 1 ∣ β γ \Psi_{>1|\beta_\gamma} Ψ > 1∣ β γ is a clause of Ψ > 1 ∣ β h \Psi_{>1|\beta_h} Ψ > 1∣ β h , and Ψ > 1 ∣ β γ = Ψ ∣ γ \Psi_{>1|\beta_\gamma} = \Psi|_\gamma Ψ > 1∣ β γ = Ψ ∣ γ by Lemma 13 . If α h \alpha_h α h and γ \gamma γ satisfy the same clauses, then β h = β γ \beta_h = \beta_\gamma β h = β γ , because both assignments are canonical.
Proposition 18. From a C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE refutation κ \kappa κ of a QCNF Φ \Phi Φ we can efficiently extract a universal winning strategy s s s such that ∣ Image ( s U j ) ∣ ≤ ∣ { S ∈ κ : Block ( S ) = U j } ∣ ≤ ∣ κ ∣ |\operatorname{Image}(s_{U_j})| \leq|\{S \in\kappa: \operatorname{Block}(S) = U_j\}| \leq|\kappa| ∣ Image ( s U j ) ∣ ≤ ∣ { S ∈ κ : Block ( S ) = U j } ∣ ≤ ∣ κ ∣ for all j = 0 , … , m j = 0,\ldots,m j = 0 , … , m with U j ≠ ∅ U_j \ne\varnothing U j = ∅ (if U 0 U_0 U 0 is empty, we ignore the trivial function s U 0 s_{U_0} s U 0 ). By summing over all universal blocks, we even obtain ∣ s ∣ ≤ ∣ κ ∣ |s| \leq|\kappa| ∣ s ∣ ≤ ∣ κ ∣ .
Proof. For every refutation slice S ∈ κ S \in\kappa S ∈ κ we define a universal winning strategy t S t^S t S for the formula Ψ \Psi Ψ of S S S , by induction along the references of κ \kappa κ . Recall that by Definition 9 the last slice referred to by a refutation slice of a universal block is a refutation slice, and all slices referred to by a refutation slice of an existential block are refutation slices.
If S = ( Ψ , π ) S = (\Psi,\pi) S = ( Ψ , π ) , then Ψ \Psi Ψ is an unsatisfiable propositional formula, and t S t^S t S is the empty strategy.
Let S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) S = (\Psi,(\alpha_1,\ldots,\alpha_g),(\beta_1,\ldots,\beta_g),(S_1,\ldots,S_g),\rho) S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) with Block ( S ) = ∀ V \operatorname{Block}(S) = \forall V Block ( S ) = ∀ V . Then S g S_g S g is a refutation slice for Ψ > 1 ∣ β g \Psi_{>1|\beta_g} Ψ > 1∣ β g , and we have Ψ > 1 ∣ β g = Ψ ∣ α g \Psi_{>1|\beta_g} = \Psi|_{\alpha_g} Ψ > 1∣ β g = Ψ ∣ α g by Lemma 13 . The strategy t S t^S t S plays α g \alpha_g α g and then follows t S g t^{S_g} t S g .
Let S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) S = (\Psi,(\alpha_1,\ldots,\alpha_g),(\beta_1,\ldots,\beta_g),(S_1,\ldots,S_g),\rho) S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) with Block ( S ) = ∃ V \operatorname{Block}(S) = \exists V Block ( S ) = ∃ V . For an assignment γ ∈ ⟨ V ⟩ \gamma\in\langle V\rangle γ ∈ ⟨ V ⟩ of the existential player let β γ \beta_\gamma β γ be canonical for γ \gamma γ with respect to Ψ \Psi Ψ , and let h h h be minimal with β h − 1 ( 1 ) ⊆ β γ − 1 ( 1 ) \beta_h^{-1}(1) \subseteq\beta_\gamma^{-1}(1) β h − 1 ( 1 ) ⊆ β γ − 1 ( 1 ) . Such an h h h exists by Lemma 17 , and we have Ψ > 1 ∣ β h ⊆ Ψ ∣ γ \Psi_{>1|\beta_h} \subseteq\Psi|_\gamma Ψ > 1∣ β h ⊆ Ψ ∣ γ , so that the strategy t S t^S t S follows t S h t^{S_h} t S h . This strategy is winning for Ψ > 1 ∣ β h \Psi_{>1|\beta_h} Ψ > 1∣ β h and therefore also for Ψ ∣ γ \Psi|_\gamma Ψ ∣ γ , since an assignment that falsifies a clause of Ψ > 1 ∣ β h \Psi_{>1|\beta_h} Ψ > 1∣ β h falsifies the matrix of Ψ ∣ γ \Psi|_\gamma Ψ ∣ γ .
By induction, t S t^S t S is a universal winning strategy for the formula of S S S . In particular, for the last slice S e S_e S e of κ \kappa κ we obtain a universal winning strategy s : = t S e s := t^{S_e} s := t S e for Φ \Phi Φ , and Φ \Phi Φ is indeed false.
For each j = 0 , … , m j = 0,\ldots,m j = 0 , … , m , every value of s U j s_{U_j} s U j is the last universal assignment α g \alpha_g α g of a slice S ∈ κ S \in\kappa S ∈ κ with Block ( S ) = U j \operatorname{Block}(S) = U_j Block ( S ) = U j . This gives ∣ Image ( s U j ) ∣ ≤ ∣ { S ∈ κ : Block ( S ) = U j } ∣ |\operatorname{Image}(s_{U_j})| \leq|\{S \in\kappa: \operatorname{Block}(S) = U_j\}| ∣ Image ( s U j ) ∣ ≤ ∣ { S ∈ κ : Block ( S ) = U j } ∣ . Finally, the construction is efficient: given κ \kappa κ and γ ∈ ⟨ X ⟩ \gamma\in\langle X\rangle γ ∈ ⟨ X ⟩ , we only have to follow one path of slices through κ \kappa κ in order to compute s ( γ ) s(\gamma) s ( γ ) .
For true QCNFs and C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE verifications, an analogous result holds.
Proposition 19. From a C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE verification κ \kappa κ of a QCNF Φ \Phi Φ we can efficiently extract an existential winning strategy s s s such that ∣ Image ( s X j ) ∣ ≤ ∣ { S ∈ κ : Block ( S ) = X j } ∣ ≤ ∣ κ ∣ |\operatorname{Image}(s_{X_j})| \leq|\{S \in\kappa: \operatorname{Block}(S) = X_j\}| \leq|\kappa| ∣ Image ( s X j ) ∣ ≤ ∣ { S ∈ κ : Block ( S ) = X j } ∣ ≤ ∣ κ ∣ for j = 1 , … , m + 1 j = 1,\ldots,m+1 j = 1 , … , m + 1 . By summing over all existential blocks, we even obtain ∣ s ∣ ≤ ∣ κ ∣ |s| \leq|\kappa| ∣ s ∣ ≤ ∣ κ ∣ .
Proof. The proof is dual to the proof of Proposition 18 . For every verification slice S ∈ κ S \in\kappa S ∈ κ we define an existential winning strategy t S t^{S} t S for the formula Ψ \Psi Ψ of S S S . By Definition 9 , the last slice referred to by a verification slice of an existential block is a verification slice, and all slices referred to by a verification slice of a universal block are verification slices.
If S = ( Ψ , σ ) S = (\Psi, \sigma) S = ( Ψ , σ ) , then t S t^{S} t S plays σ \sigma σ , which we extend arbitrarily to an assignment of all variables of X m + 1 X_{m+1} X m + 1 . This assignment satisfies Ψ \Psi Ψ .
Let S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) S = (\Psi,(\alpha_{1},\ldots,\alpha_{g}),(\beta_{1},\ldots,\beta_{g}),(S_{1},\ldots,S_{g}),\rho) S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) with Block ( S ) = ∃ V \operatorname{Block}(S) = \exists V Block ( S ) = ∃ V . Then S g S_{g} S g is a verification slice for Ψ > 1 ∣ β g \Psi_{>1|\beta_{g}} Ψ > 1∣ β g , and we have Ψ > 1 ∣ β g ⊇ Ψ ∣ α g \Psi_{>1|\beta_{g}} \supseteq\Psi_{|\alpha_{g}} Ψ > 1∣ β g ⊇ Ψ ∣ α g by Lemma 14 . The strategy t S t^{S} t S plays α g \alpha_{g} α g and then follows t S g t^{S_{g}} t S g , which satisfies all clauses of Ψ > 1 ∣ β g \Psi_{>1|\beta_{g}} Ψ > 1∣ β g and therefore all clauses of Ψ ∣ α g \Psi_{|\alpha_{g}} Ψ ∣ α g .
Let S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) S = (\Psi,(\alpha_{1},\ldots,\alpha_{g}),(\beta_{1},\ldots,\beta_{g}),(S_{1},\ldots,S_{g}),\rho) S = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) with Block ( S ) = ∀ V \operatorname{Block}(S) = \forall V Block ( S ) = ∀ V . For an assignment γ ∈ ⟨ V ⟩ \gamma\in\langle V\rangle γ ∈ ⟨ V ⟩ of the universal player let β γ \beta_{\gamma} β γ be canonical for γ \gamma γ with respect to Ψ \Psi Ψ , and let h h h be minimal with β h − 1 ( 0 ) ⊆ β γ − 1 ( 0 ) \beta_{h}^{-1}(0) \subseteq\beta_{\gamma}^{-1}(0) β h − 1 ( 0 ) ⊆ β γ − 1 ( 0 ) . Such an h h h exists by Lemma 17 , and we have Ψ > 1 ∣ β h ⊇ Ψ ∣ γ \Psi_{>1|\beta_{h}} \supseteq\Psi_{|\gamma} Ψ > 1∣ β h ⊇ Ψ ∣ γ . After γ \gamma γ , the strategy t S t^{S} t S follows t S h t^{S_{h}} t S h , which satisfies all clauses of Ψ > 1 ∣ β h \Psi_{>1|\beta_{h}} Ψ > 1∣ β h and therefore all clauses of Ψ ∣ γ \Psi_{|\gamma} Ψ ∣ γ .
By induction, t S t^{S} t S is an existential winning strategy for the formula of S S S . In particular, s : = t S e s := t^{S_{e}} s := t S e is an existential winning strategy for Φ \Phi Φ , and Φ \Phi Φ is indeed true.
For each j = 1 , … , m + 1 j = 1,\ldots,m+1 j = 1 , … , m + 1 , every value of s X j s_{X_{j}} s X j is read off a slice S ∈ κ S \in\kappa S ∈ κ with Block ( S ) = X j \operatorname{Block}(S) = X_{j} Block ( S ) = X j : it is the last existential assignment α g \alpha_{g} α g of S S S for j ≤ m j \le m j ≤ m , and the assignment σ \sigma σ for j = m + 1 j = m+1 j = m + 1 . Therefore, ∣ Image ( s X j ) ∣ ≤ ∣ { S ∈ κ : Block ( S ) = X j } ∣ |\operatorname{Image}(s_{X_{j}})| \le|\{S \in\kappa: \operatorname{Block}(S) = X_{j}\}| ∣ Image ( s X j ) ∣ ≤ ∣ { S ∈ κ : Block ( S ) = X j } ∣ . As before, the construction is efficient.
As a corollary we obtain the soundness of the proof system. Completeness will follow from the upper bounds in the next section (Corollary 31 ).
Corollary 20. C A Q E C O R E \mathit{CAQE}^{\mathit{CORE}} CAQE CORE is sound on false and on true QCNFs. In other words, if a QCNF Φ \Phi Φ has a C A Q E C O R E \mathit{CAQE}^{\mathit{CORE}} CAQE CORE refutation, then Φ \Phi Φ is false, and if Φ \Phi Φ has a C A Q E C O R E \mathit{CAQE}^{\mathit{CORE}} CAQE CORE verification, then Φ \Phi Φ is true.
Lower and upper bounds for proof size via CAQE width
Next we want to prove both lower and upper bounds for the size of C A Q E C O R E \mathit{CAQE}^{\mathit{CORE}} CAQE CORE proofs, for which we introduce a measure called C A Q E CAQE C A QE width. The main idea is as follows: We already know that CAQE only tests assignments if they are not worse than previously tested ones. This means that we can minimise the number of assignments to be tested by only picking optimal ones. Hence, as an upper bound we can count the number of these optimal assignments. On the other hand, since we need to test at least all optimal assignments, this also yields a lower bound.
The intricate part is to generalise the notion of optimal assignments to formulas with more than one quantifier alternation, as we can only check optimality on the first block. To do this, we define optimal moves on the first variable block, and optimal chains as a sequence of optimal moves for several quantifier alternations.
Definition 21 (optimal moves). Let Φ \Phi Φ be a QCNF with matrix C 1 ∧ … ∧ C k C_{1} \land\ldots\land C_{k} C 1 ∧ … ∧ C k such that its prefix starts with the quantifier block Q V QV Q V . In case Q = ∃ Q = \exists Q = ∃ , an assignment α ∈ ⟨ V ⟩ \alpha\in\langle V\rangle α ∈ ⟨ V ⟩ is called an optimal move if and only if the following holds: For each α ′ ∈ ⟨ V ⟩ \alpha' \in\langle V\rangle α ′ ∈ ⟨ V ⟩ and each clause C r C_{r} C r with α ⊭ C r \alpha\not\models C_{r} α ⊨ C r and α ′ ⊨ C r \alpha' \models C_{r} α ′ ⊨ C r there exists a clause C q C_{q} C q with α ⊨ C q \alpha\models C_{q} α ⊨ C q and α ′ ⊭ C q \alpha' \not\models C_{q} α ′ ⊨ C q . In other words, there is no assignment α ′ ∈ ⟨ V ⟩ \alpha' \in\langle V\rangle α ′ ∈ ⟨ V ⟩ such that α ′ \alpha' α ′ satisfies all clauses satisfied by α \alpha α , plus at least one more clause.
Analogously, if Q = ∀ Q = \forall Q = ∀ an assignment α ∈ ⟨ V ⟩ \alpha\in\langle V\rangle α ∈ ⟨ V ⟩ is called an optimal move if and only if the following holds: For each α ′ ∈ ⟨ V ⟩ \alpha' \in\langle V\rangle α ′ ∈ ⟨ V ⟩ and each clause C r C_r C r with α ⊨ C r \alpha\models C_r α ⊨ C r and α ′ ⊭ C r \alpha' \not\models C_r α ′ ⊨ C r there is a clause C q C_q C q with α ⊭ C q \alpha\not\models C_q α ⊨ C q and α ′ ⊨ C q \alpha' \models C_q α ′ ⊨ C q . In other words, there is no assignment α ′ ∈ ⟨ V ⟩ \alpha' \in\langle V\rangle α ′ ∈ ⟨ V ⟩ such that α \alpha α satisfies all clauses that are satisfied by α ′ \alpha' α ′ , plus at least one more clause.
For a (partial) assignment τ \tau τ of the variables of Φ \Phi Φ we denote by sat ( τ ) : = { r ∈ { 1 , … , k } : τ ⊨ C r } \operatorname{sat}(\tau) := \{r \in\{1,\ldots,k\} : \tau\models C_r\} sat ( τ ) := { r ∈ { 1 , … , k } : τ ⊨ C r } the set of indices of the clauses of Φ \Phi Φ that are satisfied by τ \tau τ . We will use this notation throughout the rest of the paper, always with respect to the original clauses C 1 , … , C k C_1,\ldots,C_k C 1 , … , C k of the QCNF Φ \Phi Φ under consideration. In the existential case, an assignment α ∈ ⟨ V ⟩ \alpha\in\langle V\rangle α ∈ ⟨ V ⟩ is then an optimal move if and only if sat ( α ) \operatorname{sat}(\alpha) sat ( α ) is a ⊆ \subseteq ⊆ -maximal element of { sat ( α ′ ) : α ′ ∈ ⟨ V ⟩ } \{\operatorname{sat}(\alpha') : \alpha' \in\langle V\rangle\} { sat ( α ′ ) : α ′ ∈ ⟨ V ⟩} , and in the universal case this holds if and only if sat ( α ) \operatorname{sat}(\alpha) sat ( α ) is a ⊆ \subseteq ⊆ -minimal element of this set. Every assignment α ′ ∈ ⟨ V ⟩ \alpha' \in\langle V\rangle α ′ ∈ ⟨ V ⟩ is dominated by an optimal move, i.e. in the existential case there is an optimal move α \alpha α with sat ( α ′ ) ⊆ sat ( α ) \operatorname{sat}(\alpha') \subseteq\operatorname{sat}(\alpha) sat ( α ′ ) ⊆ sat ( α ) , and dually in the universal case.
Optimal moves can now be generalised to optimal chains. An important aspect of optimal chains is that they always rely on a given winning strategy s s s . This is because we need to consider quantifier alternations and we need to know how the variables are assigned that are quantified between the blocks on which we want to detect optimal moves. Here we use the game notation from Definition 1.
Definition 22 (optimal chains). Let Φ \Phi Φ be a false QCNF and s s s be a universal winning strategy for Φ \Phi Φ . An assignment ( η 1 , … , η j ) ∈ ⟨ X 1 ∪ … ∪ X j ⟩ (\eta_1,\ldots,\eta_j) \in\langle X_1 \cup\ldots\cup X_j\rangle ( η 1 , … , η j ) ∈ ⟨ X 1 ∪ … ∪ X j ⟩ for j ≤ m j \le m j ≤ m is called an optimal chain up to block j j j with respect to s s s if and only if each η i ∈ ⟨ X i ⟩ \eta_i \in\langle X_i\rangle η i ∈ ⟨ X i ⟩ is an optimal move for Φ ∣ game ( s , η 1 ∪ … ∪ η i − 1 , U i − 1 ) \Phi|_{\operatorname{game}(s,\eta_1\cup\ldots\cup\eta_{i-1},U_{i-1})} Φ ∣ game ( s , η 1 ∪ … ∪ η i − 1 , U i − 1 ) for each i = 1 , … , j i=1,\ldots,j i = 1 , … , j .6
Analogously, in the case where Φ \Phi Φ is true and s s s is an existential winning strategy, an assignment ( η 0 , … , η j ) ∈ ⟨ U 0 ∪ … ∪ U j ⟩ (\eta_0,\ldots,\eta_j) \in\langle U_0 \cup\ldots\cup U_j\rangle ( η 0 , … , η j ) ∈ ⟨ U 0 ∪ … ∪ U j ⟩ for j ≤ m j \le m j ≤ m is called an optimal chain up to block j j j with respect to s s s if and only if each η i ∈ ⟨ U i ⟩ \eta_i \in\langle U_i\rangle η i ∈ ⟨ U i ⟩ is an optimal move for Φ ∣ game ( s , η 0 ∪ … ∪ η i − 1 , X i ) \Phi|_{\operatorname{game}(s,\eta_0\cup\ldots\cup\eta_{i-1},X_i)} Φ ∣ game ( s , η 0 ∪ … ∪ η i − 1 , X i ) for each i = 0 , … , j i=0,\ldots,j i = 0 , … , j .7
We denote the set of all optimal chains up to block j j j with respect to s s s as K j ( Φ , s ) K_j(\Phi,s) K j ( Φ , s ) . Furthermore, we set K ( Φ , s ) : = K m ( Φ , s ) K(\Phi,s) := K_m(\Phi,s) K ( Φ , s ) := K m ( Φ , s ) . For j ≤ 0 j \le0 j ≤ 0 (false case) and j ≤ − 1 j \le-1 j ≤ − 1 (true case) we set K j ( Φ , s ) : = { ∅ } K_j(\Phi,s) := \{\emptyset\} K j ( Φ , s ) := { ∅ } , which means that the only optimal chain up to a non-existing block is the empty assignment.
We recall the definition of maximal and minimal elements with respect to the subset relation. Let A \mathcal{A} A be a family of sets. Then we define max ⊆ ( A ) : = { M ∈ A : ∄ N ∈ A : M ⊊ N } \max_{\subseteq}(\mathcal{A}) := \{M \in\mathcal{A} : \nexists N \in\mathcal{A} : M \subsetneq N\} max ⊆ ( A ) := { M ∈ A : ∄ N ∈ A : M ⊊ N } and min ⊆ ( A ) : = { M ∈ A : ∄ N ∈ A : N ⊊ M } \min_{\subseteq}(\mathcal{A}) := \{M \in\mathcal{A} : \nexists N \in\mathcal{A} : N \subsetneq M\} min ⊆ ( A ) := { M ∈ A : ∄ N ∈ A : N ⊊ M } .
Definition 23 (CAQE width). Let Φ \Phi Φ be a QCNF. If Φ \Phi Φ is false, we define
cw − ( Φ ) : = min s max 1 ≤ j ≤ m max γ ~ ∈ K j − 1 ( Φ , s ) ∣ max ⊆ ( { r : game ( s , γ , X j ) ⊨ C r } : γ ~ ⊆ γ ∈ ⟨ X ⟩ ) ∣ \operatorname{cw}^{-}(\Phi) := \min_s \max_{1\le j\le m} \max_{\widetilde{\gamma}\in K_{j-1}(\Phi,s)} \left|\max_{\subseteq}\left(\{r : \operatorname{game}(s,\gamma,X_j) \models C_r\} : \widetilde{\gamma} \subseteq\gamma\in\langle X\rangle\right)\right| cw − ( Φ ) := s min 1 ≤ j ≤ m max γ ∈ K j − 1 ( Φ , s ) max ⊆ max ( { r : game ( s , γ , X j ) ⊨ C r } : γ ⊆ γ ∈ ⟨ X ⟩ ) and if Φ \Phi Φ is true we define
cw + ( Φ ) : = min s max 0 ≤ j ≤ m max γ ~ ∈ K j − 1 ( Φ , s ) ∣ min ⊆ ( { r : game ( s , γ , U j ) ⊨ C r } : γ ~ ⊆ γ ∈ ⟨ U ⟩ ) ∣ \operatorname{cw}^{+}(\Phi) := \min_s \max_{0\le j\le m} \max_{\widetilde{\gamma}\in K_{j-1}(\Phi,s)} \left|\min_{\subseteq}\left(\{r : \operatorname{game}(s,\gamma,U_j) \models C_r\} : \widetilde{\gamma} \subseteq\gamma\in\langle U\rangle\right)\right| cw + ( Φ ) := s min 0 ≤ j ≤ m max γ ∈ K j − 1 ( Φ , s ) max ⊆ min ( { r : game ( s , γ , U j ) ⊨ C r } : γ ⊆ γ ∈ ⟨ U ⟩ ) as the CAQE width of Φ \Phi Φ , where s s s ranges over all universal (resp. existential) winning strategies for Φ \Phi Φ .8 If m = 0 m=0 m = 0 , then the prefix of a false QCNF Φ \Phi Φ contains no existential block that is followed by a universal block, and in this case we set cw − ( Φ ) : = 1 \operatorname{cw}^{-}(\Phi) := 1 cw − ( Φ ) := 1 .
Computing the CAQE width can be interpreted as a game between two players: the strategist and the optimiser. In the case where Φ \Phi Φ is false, the strategist controls the universal variables and the optimiser controls the existential variables. The task of the strategist is to pick a universal winning strategy for Φ \Phi Φ at the beginning and assign the variables of Φ \Phi Φ along that strategy throughout the whole game. On the other hand, before assigning some block X j X_j X j the optimiser counts the number of optimal moves on that block under the side condition that different moves that have the same effect9 on the formula only count once. Let p j p_j p j be that number. Then the optimiser chooses an optimal move to play it on the formula. The goal of the optimiser is to maximise max j p j \max_j p_j max j p j by assigning X X X suitably, while the goal of the strategist is to minimise max j p j \max_j p_j max j p j by choosing the best strategy s s s at the beginning. If both players play optimally, we end up with max j p j = cw − ( Φ ) \max_j p_j=\operatorname{cw}^{-}(\Phi) max j p j = cw − ( Φ ) .
The case where Φ \Phi Φ is true is similar. There the strategist controls the existential variables and picks an existential winning strategy, whereas the optimiser assigns the universal variables. At the end we obtain max j p j = cw + ( Φ ) \max_j p_j=\operatorname{cw}^{+}(\Phi) max j p j = cw + ( Φ ) .
In the proofs we will mostly use the CAQE width in the following local form. Let Ψ \Psi Ψ be a QCNF that we obtain from Φ \Phi Φ by a restriction, with clauses indexed by a set J ⊆ { 1 , … , k } J\subseteq\{1,\ldots,k\} J ⊆ { 1 , … , k } and a prefix that starts with the block Q V QV Q V . For α ∈ ⟨ V ⟩ \alpha\in\langle V\rangle α ∈ ⟨ V ⟩ we denote by sat Ψ ( α ) ⊆ J \operatorname{sat}_{\Psi}(\alpha)\subseteq J sat Ψ ( α ) ⊆ J the set of indices of the clauses of Ψ \Psi Ψ that are satisfied by α \alpha α , and we define
w ( Ψ ) : = { ∣ max ⊆ { sat Ψ ( α ) : α ∈ ⟨ V ⟩ } ∣ if Q = ∃ , ∣ min ⊆ { sat Ψ ( α ) : α ∈ ⟨ V ⟩ } ∣ if Q = ∀ , w(\Psi):=
\begin{cases}
\left|\max_{\subseteq}\{\operatorname{sat}_{\Psi}(\alpha):\alpha\in\langle V\rangle\}\right| & \text{if }Q=\exists,\\
\left|\min_{\subseteq}\{\operatorname{sat}_{\Psi}(\alpha):\alpha\in\langle V\rangle\}\right| & \text{if }Q=\forall,
\end{cases} w ( Ψ ) := { ∣ max ⊆ { sat Ψ ( α ) : α ∈ ⟨ V ⟩} ∣ ∣ min ⊆ { sat Ψ ( α ) : α ∈ ⟨ V ⟩} ∣ if Q = ∃ , if Q = ∀ ,
which is the number of optimal moves of Ψ \Psi Ψ up to their effect on the clauses. Now let Φ \Phi Φ be false, let s s s be a universal winning strategy, γ ~ ∈ K j − 1 ( Φ , s ) \widetilde{\gamma}\in K_{j-1}(\Phi,s) γ ∈ K j − 1 ( Φ , s ) and Ψ : = Φ ∣ game ( s , γ ~ , U ~ j − 1 ) \Psi:=\Phi|_{\operatorname{game}(s,\widetilde{\gamma},\widetilde{U}_{j-1})} Ψ := Φ ∣ game ( s , γ , U j − 1 ) . For every γ ∈ ⟨ X ⟩ \gamma\in\langle X\rangle γ ∈ ⟨ X ⟩ with γ ~ ⊆ γ \widetilde{\gamma}\subseteq\gamma γ ⊆ γ we have game ( s , γ , X j ) = game ( s , γ ~ , U ~ j − 1 ) ∪ γ j \operatorname{game}(s,\gamma,X_j)=\operatorname{game}(s,\widetilde{\gamma},\widetilde{U}_{j-1})\cup\gamma_j game ( s , γ , X j ) = game ( s , γ , U j − 1 ) ∪ γ j . Hence, the sets { r : game ( s , γ , X j ) ⊨ C r } \{r:\operatorname{game}(s,\gamma,X_j)\models C_r\} { r : game ( s , γ , X j ) ⊨ C r } with γ ~ ⊆ γ \widetilde{\gamma}\subseteq\gamma γ ⊆ γ are exactly the sets I ∪ sat Ψ ( α ) I\cup\operatorname{sat}_{\Psi}(\alpha) I ∪ sat Ψ ( α ) for α ∈ ⟨ X j ⟩ \alpha\in\langle X_j\rangle α ∈ ⟨ X j ⟩ , where I I I is the fixed set of clauses satisfied by game ( s , γ ~ , U ~ j − 1 ) \operatorname{game}(s,\widetilde{\gamma},\widetilde{U}_{j-1}) game ( s , γ , U j − 1 ) . As a consequence, the number of ⊆ \subseteq ⊆ -maximal sets in Definition 23 is w ( Ψ ) w(\Psi) w ( Ψ ) , and we get
cw − ( Φ ) = min s max 1 ≤ j ≤ m max γ ~ ∈ K j − 1 ( Φ , s ) w ( Φ ∣ game ( s , γ ~ , U ~ j − 1 ) ) . (1) \operatorname{cw}^{-}(\Phi)=\min_s\max_{1\leq j\leq m}\max_{\widetilde{\gamma}\in K_{j-1}(\Phi,s)}w\left(\Phi|_{\operatorname{game}(s,\widetilde{\gamma},\widetilde{U}_{j-1})}\right). \tag*{(1)} cw − ( Φ ) = s min 1 ≤ j ≤ m max γ ∈ K j − 1 ( Φ , s ) max w ( Φ ∣ game ( s , γ , U j − 1 ) ) . (1)
In the same way we obtain cw + ( Φ ) = min s max 0 ≤ j ≤ m max γ ~ ∈ K j − 1 ( Φ , s ) w ( Φ ∣ game ( s , γ ~ , X ~ j ) ) \operatorname{cw}^{+}(\Phi)=\min_s\max_{0\leq j\leq m}\max_{\widetilde{\gamma}\in K_{j-1}(\Phi,s)}w\left(\Phi|_{\operatorname{game}(s,\widetilde{\gamma},\widetilde{X}_j)}\right) cw + ( Φ ) = min s max 0 ≤ j ≤ m max γ ∈ K j − 1 ( Φ , s ) w ( Φ ∣ game ( s , γ , X j ) ) for a true QCNF Φ \Phi Φ .
Example 24. Let Φ \Phi Φ be the QCNF from Example 2. The first block of Φ \Phi Φ is existential and m = 1 m=1 m = 1 , so the CAQE width of Φ \Phi Φ does not depend on the strategy s s s , and cw − ( Φ ) \operatorname{cw}^{-}(\Phi) cw − ( Φ ) is simply the number of ⊆ \subseteq ⊆ -maximal sets among sat ( α ) \operatorname{sat}(\alpha) sat ( α ) for α ∈ ⟨ x 1 , x 2 ⟩ \alpha\in\langle x_1,x_2\rangle α ∈ ⟨ x 1 , x 2 ⟩ . We already computed these sets in Example 2: they are sat ( α 1 ) = { 4 , 5 , 6 } \operatorname{sat}(\alpha_1)=\{4,5,6\} sat ( α 1 ) = { 4 , 5 , 6 } , sat ( α 2 ) = { 2 , 4 , 5 } \operatorname{sat}(\alpha_2)=\{2,4,5\} sat ( α 2 ) = { 2 , 4 , 5 } and sat ( α 3 ) = { 1 , 2 , 6 } \operatorname{sat}(\alpha_3)=\{1,2,6\} sat ( α 3 ) = { 1 , 2 , 6 } , whereas sat ( α 4 ) = { 1 , 2 } ⊆ sat ( α 3 ) \operatorname{sat}(\alpha_4)=\{1,2\}\subseteq\operatorname{sat}(\alpha_3) sat ( α 4 ) = { 1 , 2 } ⊆ sat ( α 3 ) . This gives cw − ( Φ ) = 3 \operatorname{cw}^{-}(\Phi)=3 cw − ( Φ ) = 3 .
There are formulas for which we need to count all assignments for CAQE width. Examples are the equality formulas, in which each assignment of the first block is optimal by symmetry.
Definition 25 (equality formulas [4 , 19 ] ). The false QCNF E q n \mathrm{Eq}_n Eq n consists of the prefix ∃ x 1 , … , x n ∀ u 1 , … , u n ∃ t 1 , … , t n \exists x_1,\ldots,x_n\ \forall u_1,\ldots,u_n\exists t_1,\ldots,t_n ∃ x 1 , … , x n ∀ u 1 , … , u n ∃ t 1 , … , t n and the matrix x i ∨ u i ∨ t i , x ˉ i ∨ u ˉ i ∨ t i , t ˉ 1 ∨ ⋯ ∨ t ˉ n x_i\lor u_i\lor t_i,\ \bar{x}_i\lor\bar{u}_i\lor t_i,\ \bar{t}_1\lor\cdots\lor\bar{t}_n x i ∨ u i ∨ t i , x ˉ i ∨ u ˉ i ∨ t i , t ˉ 1 ∨ ⋯ ∨ t ˉ n for i = 1 , … , n i=1,\ldots,n i = 1 , … , n .
We use the form of the equality formulas from [19 , 18 ] . The original definition in [4 ] is the same up to renaming t i ↦ t ˉ i t_i\mapsto\bar{t}_i t i ↦ t ˉ i .
Proposition 26. We have cw − ( E q n ) = 2 n \operatorname{cw}^{-}(\mathrm{Eq}_{n}) = 2^{n} cw − ( Eq n ) = 2 n .
Proof. Let γ 1 , γ 2 ∈ ⟨ { x 1 , … , x n } ⟩ \gamma_{1},\gamma_{2} \in\langle\{x_{1},\ldots,x_{n}\}\rangle γ 1 , γ 2 ∈ ⟨{ x 1 , … , x n }⟩ be two assignments with γ 1 ≠ γ 2 \gamma_{1} \ne\gamma_{2} γ 1 = γ 2 . Then there exists an i i i with γ 1 ( x i ) ≠ γ 2 ( x i ) \gamma_{1}(x_{i}) \ne\gamma_{2}(x_{i}) γ 1 ( x i ) = γ 2 ( x i ) . Hence, x i ∨ u i ∨ t i x_{i} \lor u_{i} \lor t_{i} x i ∨ u i ∨ t i is satisfied by γ 1 \gamma_{1} γ 1 but not by γ 2 \gamma_{2} γ 2 , and x ‾ i ∨ u ‾ i ∨ t i \overline{x}_{i} \lor\overline{u}_{i} \lor t_{i} x i ∨ u i ∨ t i is satisfied by γ 2 \gamma_{2} γ 2 but not by γ 1 \gamma_{1} γ 1 , or vice versa. That means all assignments of the first block are optimal. The first block is existential and m = 1 m = 1 m = 1 , hence the CAQE width does not depend on the strategy and we obtain cw − ( E q n ) = ∣ ⟨ { x 1 , … , x n } ⟩ ∣ = 2 n \operatorname{cw}^{-}(\mathrm{Eq}_{n}) = |\langle\{x_{1},\ldots,x_{n}\}\rangle| = 2^{n} cw − ( Eq n ) = ∣ ⟨{ x 1 , … , x n }⟩ ∣ = 2 n .
Next we define fruitions . These are projections of assignments that tell us which clauses of a formula Φ \Phi Φ are satisfied by these assignments. Like optimal chains, they are defined across several quantifier alternations and hence need a fixed winning strategy s s s as input. Intuitively, while assignments are purely syntactic, their fruitions encode the effect of their assignments on the formula. We will need fruitions for the decision lists in Sections 7 and § 8 .
Definition 27 (fruitions). Let Φ \Phi Φ be a QCNF.
If Φ \Phi Φ is false and s s s is a universal winning strategy, then for each assignment σ = ( σ 1 , … , σ j ) ∈ ⟨ X 1 ∪ … ∪ X j ⟩ \sigma= (\sigma_{1},\ldots,\sigma_{j}) \in\langle X_{1} \cup\ldots\cup X_{j}\rangle σ = ( σ 1 , … , σ j ) ∈ ⟨ X 1 ∪ … ∪ X j ⟩ with j ≤ m j \le m j ≤ m we call fru ( σ ) : = ( I 1 , … , I j ) \operatorname{fru}(\sigma) := (I_{1},\ldots,I_{j}) fru ( σ ) := ( I 1 , … , I j ) the fruition of σ \sigma σ with I i : = { r : game ( s , σ , X i ) ⊨ C r } I_{i} := \{r : \operatorname{game}(s,\sigma,X_{i}) \models C_{r}\} I i := { r : game ( s , σ , X i ) ⊨ C r } , for i = 1 , … , j i = 1,\ldots,j i = 1 , … , j .
Analogously, if Φ \Phi Φ is true and s s s is an existential winning strategy, then for each assignment σ = ( σ 0 , … , σ j ) ∈ ⟨ U 0 ∪ … ∪ U j ⟩ \sigma= (\sigma_{0},\ldots,\sigma_{j}) \in\langle U_{0} \cup\ldots\cup U_{j}\rangle σ = ( σ 0 , … , σ j ) ∈ ⟨ U 0 ∪ … ∪ U j ⟩ with j ≤ m j \le m j ≤ m we call fru ( σ ) : = ( I 0 , … , I j ) \operatorname{fru}(\sigma) := (I_{0},\ldots,I_{j}) fru ( σ ) := ( I 0 , … , I j ) the fruition of σ \sigma σ with I i : = { r : game ( s , σ , U i ) ⊨ C r } I_{i} := \{r : \operatorname{game}(s,\sigma,U_{i}) \models C_{r}\} I i := { r : game ( s , σ , U i ) ⊨ C r } , for i = 0 , … , j i = 0,\ldots,j i = 0 , … , j .
We denote F j ( Φ , s ) : = { fru ( σ ) : σ ∈ K j ( Φ , s ) } F_{j}(\Phi,s) := \{\operatorname{fru}(\sigma) : \sigma\in K_{j}(\Phi,s)\} F j ( Φ , s ) := { fru ( σ ) : σ ∈ K j ( Φ , s )} and set F ( Φ , s ) : = F m ( Φ , s ) F(\Phi,s) := F_{m}(\Phi,s) F ( Φ , s ) := F m ( Φ , s ) . We will sometimes refer to the elements of F j ( Φ , s ) F_{j}(\Phi,s) F j ( Φ , s ) as optimal fruitions (up to block j j j with respect to s s s ). Note that I 1 ⊆ I 2 ⊆ … I_{1} \subseteq I_{2} \subseteq\ldots I 1 ⊆ I 2 ⊆ … for every fruition, since assignments are only extended along the game.
We can now prove an upper bound for C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proof size using CAQE width. The main idea is to construct C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proofs by always choosing optimal moves for testing, and their number is bounded by the CAQE width. As a side result, we also prove the completeness of C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core by showing how refutations or verifications can always be constructed.
The following lemma shows that the optimal moves of a formula, one for each effect, always form a slice.
Lemma 28. Let Ψ \Psi Ψ be a QCNF that is not propositional and whose prefix starts with the block ∃ V \exists V ∃ V . Let α 1 , … , α g ∈ ⟨ V ⟩ \alpha_{1},\ldots,\alpha_{g} \in\langle V\rangle α 1 , … , α g ∈ ⟨ V ⟩ be such that the sets sat Ψ ( α 1 ) , … , sat Ψ ( α g ) \operatorname{sat}_{\Psi}(\alpha_{1}),\ldots,\operatorname{sat}_{\Psi}(\alpha_{g}) sat Ψ ( α 1 ) , … , sat Ψ ( α g ) are pairwise different and are exactly the ⊆ \subseteq ⊆ -maximal elements of { sat Ψ ( α ) : α ∈ ⟨ V ⟩ } \{\operatorname{sat}_{\Psi}(\alpha) : \alpha\in\langle V\rangle\} { sat Ψ ( α ) : α ∈ ⟨ V ⟩} , and let β i \beta_{i} β i be canonical for α i \alpha_{i} α i . Then (1) α i ∪ β i \alpha_{i} \cup\beta_{i} α i ∪ β i satisfies Ψ ∃ V ∧ ⋀ a ≠ i G β ‾ a ∃ \Psi_{\exists V} \land\bigwedge_{a \ne i} G^{\exists}_{\overline{\beta}_{a}} Ψ ∃ V ∧ ⋀ a = i G β a ∃ for i = 1 , … , g i = 1,\ldots,g i = 1 , … , g , (2) Ψ ∃ V ∧ ⋀ a = 1 g G β ‾ a ∃ \Psi_{\exists V} \land\bigwedge_{a=1}^{g} G^{\exists}_{\overline{\beta}_{a}} Ψ ∃ V ∧ ⋀ a = 1 g G β a ∃ is unsatisfiable, and (3) Ψ > 1 ∣ β i = Ψ ∣ α i \Psi_{>1|\beta_{i}} = \Psi_{|\alpha_{i}} Ψ > 1∣ β i = Ψ ∣ α i for i = 1 , … , g i = 1,\ldots,g i = 1 , … , g .
Hence, if S i S_{i} S i is a refutation slice for Ψ ∣ α i \Psi_{|\alpha_{i}} Ψ ∣ α i for i = 1 , … , g i = 1,\ldots,g i = 1 , … , g and ρ \rho ρ is a Res refutation of the formula in (2), then ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) (\Psi,(\alpha_{1},\ldots,\alpha_{g}),(\beta_{1},\ldots,\beta_{g}),(S_{1},\ldots,S_{g}),\rho) ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S 1 , … , S g ) , ρ ) is a refutation slice for Ψ \Psi Ψ of size g = w ( Ψ ) g = w(\Psi) g = w ( Ψ ) .
The dual statement holds if the prefix of Ψ \Psi Ψ starts with ∀ V \forall V ∀ V . Here we replace ⊆ \subseteq ⊆ -maximal by ⊆ \subseteq ⊆ -minimal, Ψ ∃ V \Psi_{\exists V} Ψ ∃ V by Ψ ∀ V \Psi_{\forall V} Ψ ∀ V , G β ‾ a ∃ G^{\exists}_{\overline{\beta}_{a}} G β a ∃ by G β ‾ a ∀ G^{\forall}_{\overline{\beta}_{a}} G β a ∀ , and refutation slices by verification slices.
Proof. By Lemma 5, α i ∪ β i \alpha_{i} \cup\beta_{i} α i ∪ β i satisfies Ψ ∃ V \Psi_{\exists V} Ψ ∃ V . For a ≠ i a \ne i a = i the sets sat Ψ ( α i ) \operatorname{sat}_{\Psi}(\alpha_{i}) sat Ψ ( α i ) and sat Ψ ( α a ) \operatorname{sat}_{\Psi}(\alpha_{a}) sat Ψ ( α a ) are different and ⊆ \subseteq ⊆ -maximal, so there is some r ∈ sat Ψ ( α i ) ∖ sat Ψ ( α a ) r \in\operatorname{sat}_{\Psi}(\alpha_{i}) \setminus\operatorname{sat}_{\Psi}(\alpha_{a}) r ∈ sat Ψ ( α i ) ∖ sat Ψ ( α a ) . Then β i ( b r ) = 0 \beta_{i}(b_{r}) = 0 β i ( b r ) = 0 and β a ( b r ) = 1 \beta_{a}(b_{r}) = 1 β a ( b r ) = 1 , hence β i \beta_{i} β i satisfies the literal b ‾ r \overline{b}_{r} b r of G β ‾ a ∃ G^{\exists}_{\overline{\beta}_{a}} G β a ∃ . This shows (1).
For (2), let α ~ ∪ β ~ \widetilde{\alpha} \cup\widetilde{\beta} α ∪ β be an arbitrary assignment that satisfies Ψ ∃ V \Psi_{\exists V} Ψ ∃ V . We choose a a a with sat Ψ ( α ~ ) ⊆ sat Ψ ( α a ) \operatorname{sat}_{\Psi}(\widetilde{\alpha}) \subseteq\operatorname{sat}_{\Psi}(\alpha_{a}) sat Ψ ( α ) ⊆ sat Ψ ( α a ) . For every r r r with β a ( b r ) = 1 \beta_{a}(b_{r}) = 1 β a ( b r ) = 1 , the clause C r C_{r} C r is not satisfied by α a \alpha_{a} α a and therefore not by α ~ \widetilde{\alpha} α . As α ~ ∪ β ~ \widetilde{\alpha} \cup\widetilde{\beta} α ∪ β satisfies C r , 1 C_{r,1} C r , 1 , we get β ~ ( b r ) = 1 \widetilde{\beta}(b_{r}) = 1 β ( b r ) = 1 . Hence, β ~ \widetilde{\beta} β falsifies G β ‾ a ∃ G^{\exists}_{\overline{\beta}_{a}} G β a ∃ .
Statement (3) holds by Lemma 14, and (1)–(3) are exactly the conditions of Definition 9 for a refutation slice.
In the dual case we argue in the same way. For (1) we take some r ∈ sat Ψ ( α a ) ∖ sat Ψ ( α i ) r \in\operatorname{sat}_{\Psi}(\alpha_a) \setminus\operatorname{sat}_{\Psi}(\alpha_i) r ∈ sat Ψ ( α a ) ∖ sat Ψ ( α i ) , so that β i \beta_i β i satisfies the literal b r b_r b r of G β a V G^{\mathsf{V}}_{\beta_a} G β a V . For (2) we choose a a a with sat Ψ ( α a ) ⊆ sat Ψ ( α ~ ) \operatorname{sat}_{\Psi}(\alpha_a) \subseteq\operatorname{sat}_{\Psi}(\widetilde{\alpha}) sat Ψ ( α a ) ⊆ sat Ψ ( α ) . For every r r r with β a ( b r ) = 0 \beta_a(b_r)=0 β a ( b r ) = 0 the assignment α ~ \widetilde{\alpha} α satisfies some literal ℓ \ell ℓ of C r C_r C r , and the clause ℓ ‾ ∨ b r ‾ \overline{\ell} \lor\overline{b_r} ℓ ∨ b r of C r , 1 C_{r,1} C r , 1 gives us β ~ ( b r ) = 0 \widetilde{\beta}(b_r)=0 β ( b r ) = 0 . Hence, β ~ \widetilde{\beta} β falsifies G β a V G^{\mathsf{V}}_{\beta_a} G β a V . For (3) we use Lemma 13. Note that the assignments β i \beta_i β i are canonical, as Definition 9 requires for universal blocks.
Theorem 29. For each false QCNF Φ \Phi Φ there exists a CAQEC O R E ^{\mathrm{CORE}} CORE refutation of size at most ( 2 m + 2 ) ⋅ cw − ( Φ ) m (2m+2)\cdot\operatorname{cw}^{-}(\Phi)^m ( 2 m + 2 ) ⋅ cw − ( Φ ) m . In particular, Φ \Phi Φ has a CAQEC O R E ^{\mathrm{CORE}} CORE refutation of size O ( cw − ( Φ ) m ) O(\operatorname{cw}^{-}(\Phi)^m) O ( cw − ( Φ ) m ) if cw − ( Φ ) ≥ 2 \operatorname{cw}^{-}(\Phi)\ge2 cw − ( Φ ) ≥ 2 .
Proof. Let s s s be a universal winning strategy that attains the minimum in Definition 23, and let c : = cw − ( Φ ) c:=\operatorname{cw}^{-}(\Phi) c := cw − ( Φ ) . By (1), we have w ( Φ ∣ game ( s , γ ~ , U j − 1 ) ) ≤ c w(\Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_{j-1})})\le c w ( Φ ∣ game ( s , γ , U j − 1 ) ) ≤ c for all 1 ≤ j ≤ m 1\le j\le m 1 ≤ j ≤ m and all γ ~ ∈ K j − 1 ( Φ , s ) \widetilde{\gamma}\in K_{j-1}(\Phi,s) γ ∈ K j − 1 ( Φ , s ) .
We define refutation slices S X ( γ ~ ) S^{X}(\widetilde{\gamma}) S X ( γ ) for the formulas Φ ∣ game ( s , γ ~ , U j − 1 ) \Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_{j-1})} Φ ∣ game ( s , γ , U j − 1 ) with γ ~ ∈ K j − 1 ( Φ , s ) \widetilde{\gamma}\in K_{j-1}(\Phi,s) γ ∈ K j − 1 ( Φ , s ) and 1 ≤ j ≤ m + 1 1\le j\le m+1 1 ≤ j ≤ m + 1 , and refutation slices S U ( γ ~ ) S^{U}(\widetilde{\gamma}) S U ( γ ) for the formulas Φ ∣ game ( s , γ ~ , X j ) \Phi|_{\operatorname{game}(s,\widetilde{\gamma},X_j)} Φ ∣ game ( s , γ , X j ) with γ ~ ∈ K j ( Φ , s ) \widetilde{\gamma}\in K_j(\Phi,s) γ ∈ K j ( Φ , s ) and 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m . All these formulas are false, because s s s is a universal winning strategy. The definition is by recursion, starting from the last quantifier block.
Let γ ~ ∈ K m ( Φ , s ) \widetilde{\gamma}\in K_m(\Phi,s) γ ∈ K m ( Φ , s ) . Then Φ ∣ game ( s , γ ~ , U m ) \Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_m)} Φ ∣ game ( s , γ , U m ) is an unsatisfiable propositional formula, and we set S X ( γ ~ ) : = ( Φ ∣ game ( s , γ ~ , U m ) , π ) S^{X}(\widetilde{\gamma}):=(\Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_m)},\pi) S X ( γ ) := ( Φ ∣ game ( s , γ , U m ) , π ) for a Res refutation π \pi π of this formula.
Let γ ~ ∈ K j ( Φ , s ) \widetilde{\gamma}\in K_j(\Phi,s) γ ∈ K j ( Φ , s ) with 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m and Ψ : = Φ ∣ game ( s , γ ~ , X j ) \Psi:=\Phi|_{\operatorname{game}(s,\widetilde{\gamma},X_j)} Ψ := Φ ∣ game ( s , γ , X j ) (for j = 0 j=0 j = 0 only if U 0 U_0 U 0 is not empty). Let α : = s U j ( γ ~ ) \alpha:=s_{U_j}(\widetilde{\gamma}) α := s U j ( γ ) and let β \beta β be canonical for α \alpha α . Then α ∪ β \alpha\cup\beta α ∪ β satisfies Ψ ∀ U j \Psi^{\forall U_j} Ψ ∀ U j by Lemma 5, and we have Ψ > 1 ∣ β = Ψ ∣ α = Φ ∣ game ( s , γ ~ , U j ) \Psi_{>1}|_{\beta}=\Psi|_{\alpha}=\Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_j)} Ψ > 1 ∣ β = Ψ ∣ α = Φ ∣ game ( s , γ , U j ) by Lemma 13. Hence, S U ( γ ~ ) : = ( Ψ , ( α ) , ( β ) , ( S X ( γ ~ ) , ρ ) ) S^{U}(\widetilde{\gamma}):=(\Psi,(\alpha),(\beta),(S^{X}(\widetilde{\gamma}),\rho)) S U ( γ ) := ( Ψ , ( α ) , ( β ) , ( S X ( γ ) , ρ )) with a placeholder ρ \rho ρ is a refutation slice for Ψ \Psi Ψ of size 1.
Let γ ~ ∈ K j − 1 ( Φ , s ) \widetilde{\gamma}\in K_{j-1}(\Phi,s) γ ∈ K j − 1 ( Φ , s ) with 1 ≤ j ≤ m 1\le j\le m 1 ≤ j ≤ m and Ψ : = Φ ∣ game ( s , γ ~ , U j − 1 ) \Psi:=\Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_{j-1})} Ψ := Φ ∣ game ( s , γ , U j − 1 ) . We choose α 1 , … , α g ∈ ⟨ X j ⟩ \alpha_1,\ldots,\alpha_g\in\langle X_j\rangle α 1 , … , α g ∈ ⟨ X j ⟩ as in Lemma 28, that means one optimal move of Ψ \Psi Ψ for each ⊆ \subseteq ⊆ -maximal set sat Ψ ( α ) \operatorname{sat}_{\Psi}(\alpha) sat Ψ ( α ) . Then g = w ( Ψ ) ≤ c g=w(\Psi)\le c g = w ( Ψ ) ≤ c , and for each i i i we have γ ~ ∪ α i ∈ K j ( Φ , s ) \widetilde{\gamma}\cup\alpha_i\in K_j(\Phi,s) γ ∪ α i ∈ K j ( Φ , s ) and Ψ ∣ α i = Φ ∣ game ( s , γ ~ ∪ α i , X j ) \Psi|_{\alpha_i}=\Phi|_{\operatorname{game}(s,\widetilde{\gamma}\cup\alpha_i,X_j)} Ψ ∣ α i = Φ ∣ game ( s , γ ∪ α i , X j ) . By Lemma 28,
S X ( γ ~ ) : = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S U ( γ ~ ∪ α 1 ) , … , S U ( γ ~ ∪ α g ) ) , ρ ) S^{X}(\widetilde{\gamma}):=(\Psi,(\alpha_1,\ldots,\alpha_g),(\beta_1,\ldots,\beta_g),(S^{U}(\widetilde{\gamma}\cup\alpha_1),\ldots,S^{U}(\widetilde{\gamma}\cup\alpha_g)),\rho) S X ( γ ) := ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S U ( γ ∪ α 1 ) , … , S U ( γ ∪ α g )) , ρ ) with the canonical assignments β i \beta_i β i and a suitable Res refutation ρ \rho ρ is a refutation slice for Ψ \Psi Ψ of size at most c c c .
Let κ \kappa κ consist of the slice for Φ \Phi Φ – this is S U ( ∅ ) S^{U}(\varnothing) S U ( ∅ ) , or S X ( ∅ ) S^{X}(\varnothing) S X ( ∅ ) if U 0 U_0 U 0 is empty – and of all slices that can be reached from it via references, in an order such that each slice occurs after the slices it refers to. Then κ \kappa κ is a CAQEC O R E ^{\mathrm{CORE}} CORE refutation of Φ \Phi Φ .
Since each slice S X ( γ ~ ) S^{X}(\widetilde{\gamma}) S X ( γ ) refers to at most c c c slices and each slice S U ( γ ~ ) S^{U}(\widetilde{\gamma}) S U ( γ ) refers to only one slice, κ \kappa κ contains for each j = 1 , … , m j=1,\ldots,m j = 1 , … , m at most c j c^j c j slices of the block ∃ X j \exists X_j ∃ X j , each of size at most c c c , and for each j = 0 , … , m j=0,\ldots,m j = 0 , … , m at most c j c^j c j slices of the block ∀ U j \forall U_j ∀ U j , each of size 1. In addition, there are at most c m c^m c m propositional slices of size 1. We obtain ∣ κ ∣ ≤ 1 + ∑ j = 1 m ( c j + c j ) + c m ≤ ( 2 m + 2 ) ⋅ c m |\kappa|\le1+\sum_{j=1}^{m}(c^j+c^j)+c^m\le(2m+2)\cdot c^m ∣ κ ∣ ≤ 1 + ∑ j = 1 m ( c j + c j ) + c m ≤ ( 2 m + 2 ) ⋅ c m .10 If c ≥ 2 c\ge2 c ≥ 2 , then ∑ j = 0 m c j ≤ 2 c m \sum_{j=0}^{m}c^j\le2c^m ∑ j = 0 m c j ≤ 2 c m and hence ∣ κ ∣ ≤ 5 c m ∈ O ( c m ) |\kappa|\le5c^m\in O(c^m) ∣ κ ∣ ≤ 5 c m ∈ O ( c m ) .
An analogous result holds for true QCNFs.
Proposition 30. For each true QCNF Φ \Phi Φ there exists a CAQEC O R E ^{\mathrm{CORE}} CORE verification κ \kappa κ of size at most ( 2 m + 2 ) ⋅ cw + ( Φ ) m + 1 (2m+2)\cdot\operatorname{cw}^{+}(\Phi)^{m+1} ( 2 m + 2 ) ⋅ cw + ( Φ ) m + 1 . In particular, ∣ κ ∣ ∈ O ( cw + ( Φ ) m + 1 ) |\kappa|\in O(\operatorname{cw}^{+}(\Phi)^{m+1}) ∣ κ ∣ ∈ O ( cw + ( Φ ) m + 1 ) if cw + ( Φ ) ≥ 2 \operatorname{cw}^{+}(\Phi)\ge2 cw + ( Φ ) ≥ 2 .11
Proof. The proof is dual to the proof of Theorem 29 . Let s s s be an existential winning strategy that attains the minimum in Definition 23 , and let c : = cw + ( Φ ) c := \operatorname{cw}^{+}(\Phi) c := cw + ( Φ ) . Then we have w ( Φ ∣ game ( s , γ ~ , X j ) ) ≤ c w(\Phi|_{\operatorname{game}(s,\widetilde{\gamma},X_j)}) \le c w ( Φ ∣ game ( s , γ , X j ) ) ≤ c for all 0 ≤ j ≤ m 0 \le j \le m 0 ≤ j ≤ m and all γ ~ ∈ K j − 1 ( Φ , s ) \widetilde{\gamma} \in K_{j-1}(\Phi,s) γ ∈ K j − 1 ( Φ , s ) .
For γ ~ ∈ K j − 1 ( Φ , s ) \widetilde{\gamma} \in K_{j-1}(\Phi,s) γ ∈ K j − 1 ( Φ , s ) we define verification slices S X ( γ ~ ) S^{X}(\widetilde{\gamma}) S X ( γ ) for the formulas Φ ∣ game ( s , γ ~ , U j − 1 ) \Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_{j-1})} Φ ∣ game ( s , γ , U j − 1 ) with 1 ≤ j ≤ m + 1 1 \le j \le m+1 1 ≤ j ≤ m + 1 , and verification slices S U ( γ ~ ) S^{U}(\widetilde{\gamma}) S U ( γ ) for the formulas Φ ∣ game ( s , γ ~ , X j ) \Phi|_{\operatorname{game}(s,\widetilde{\gamma},X_j)} Φ ∣ game ( s , γ , X j ) with 0 ≤ j ≤ m 0 \le j \le m 0 ≤ j ≤ m . All these formulas are true, because s s s is an existential winning strategy. Again, the definition is by recursion, starting from the last quantifier block.
Let γ ~ ∈ K m ( Φ , s ) \widetilde{\gamma} \in K_m(\Phi,s) γ ∈ K m ( Φ , s ) . Then Φ ∣ game ( s , γ ~ , U m ) \Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_m)} Φ ∣ game ( s , γ , U m ) is a satisfiable propositional formula, and we set S X ( γ ~ ) : = ( Φ ∣ game ( s , γ ~ , U m ) , σ ) S^{X}(\widetilde{\gamma}) := (\Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_m)},\sigma) S X ( γ ) := ( Φ ∣ game ( s , γ , U m ) , σ ) for a satisfying assignment σ \sigma σ of this formula.
Let γ ~ ∈ K j − 1 ( Φ , s ) \widetilde{\gamma} \in K_{j-1}(\Phi,s) γ ∈ K j − 1 ( Φ , s ) with 1 ≤ j ≤ m 1 \le j \le m 1 ≤ j ≤ m and Ψ : = Φ ∣ game ( s , γ ~ , U j − 1 ) \Psi:= \Phi|_{\operatorname{game}(s,\widetilde{\gamma},U_{j-1})} Ψ := Φ ∣ game ( s , γ , U j − 1 ) . Let α : = s X j ( γ ~ ) \alpha:= s_{X_j}(\widetilde{\gamma}) α := s X j ( γ ) and let β \beta β be canonical for α \alpha α . Then α ∪ β \alpha\cup\beta α ∪ β satisfies Ψ ∃ X j \Psi_{\exists X_j} Ψ ∃ X j by Lemma 5 , and we have Ψ > 1 ∣ β = Ψ ∣ α = Φ ∣ game ( s , γ ~ , X j ) \Psi_{>1}|_{\beta} = \Psi|_{\alpha} = \Phi|_{\operatorname{game}(s,\widetilde{\gamma},X_j)} Ψ > 1 ∣ β = Ψ ∣ α = Φ ∣ game ( s , γ , X j ) by Lemma 14 . Hence, S X ( γ ~ ) : = ( Ψ , ( α ) , ( β ) , ( S U ( γ ~ ) ) , ρ ) S^{X}(\widetilde{\gamma}) := (\Psi,(\alpha),(\beta),(S^{U}(\widetilde{\gamma})),\rho) S X ( γ ) := ( Ψ , ( α ) , ( β ) , ( S U ( γ )) , ρ ) with a placeholder ρ \rho ρ is a verification slice for Ψ \Psi Ψ of size 1.
Let γ ~ ∈ K j − 1 ( Φ , s ) \widetilde{\gamma} \in K_{j-1}(\Phi,s) γ ∈ K j − 1 ( Φ , s ) with 0 ≤ j ≤ m 0 \le j \le m 0 ≤ j ≤ m and Ψ : = Φ ∣ game ( s , γ ~ , X j ) \Psi:= \Phi|_{\operatorname{game}(s,\widetilde{\gamma},X_j)} Ψ := Φ ∣ game ( s , γ , X j ) (for j = 0 j=0 j = 0 only if U 0 U_0 U 0 is not empty). We choose α 1 , … , α g ∈ ⟨ U j ⟩ \alpha_1,\ldots,\alpha_g \in\langle U_j\rangle α 1 , … , α g ∈ ⟨ U j ⟩ as in the dual case of Lemma 28 , that means one optimal move of Ψ \Psi Ψ for each ⊆ \subseteq ⊆ -minimal set sat Ψ ( α ) \operatorname{sat}_{\Psi}(\alpha) sat Ψ ( α ) . Then g = w ( Ψ ) ≤ c g = w(\Psi) \le c g = w ( Ψ ) ≤ c , and for each i i i we have γ ~ ∪ α i ∈ K j ( Φ , s ) \widetilde{\gamma}\cup\alpha_i \in K_j(\Phi,s) γ ∪ α i ∈ K j ( Φ , s ) and Ψ ∣ α i = Φ ∣ game ( s , γ ~ ∪ α i , U j ) \Psi|_{\alpha_i} = \Phi|_{\operatorname{game}(s,\widetilde{\gamma}\cup\alpha_i,U_j)} Ψ ∣ α i = Φ ∣ game ( s , γ ∪ α i , U j ) . By Lemma 28 ,
S U ( γ ~ ) : = ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S X ( γ ~ ∪ α 1 ) , … , S X ( γ ~ ∪ α g ) ) , ρ ) S^{U}(\widetilde{\gamma}) := \left(\Psi,(\alpha_1,\ldots,\alpha_g),(\beta_1,\ldots,\beta_g),\left(S^{X}(\widetilde{\gamma}\cup\alpha_1),\ldots,S^{X}(\widetilde{\gamma}\cup\alpha_g)\right),\rho\right) S U ( γ ) := ( Ψ , ( α 1 , … , α g ) , ( β 1 , … , β g ) , ( S X ( γ ∪ α 1 ) , … , S X ( γ ∪ α g ) ) , ρ )
with the canonical assignments β i \beta_i β i and a suitable Res refutation ρ \rho ρ is a verification slice for Ψ \Psi Ψ of size at most c c c .
Let κ \kappa κ consist of the slice for Φ \Phi Φ – this is S U ( ∅ ) S^{U}(\varnothing) S U ( ∅ ) , or S X ( ∅ ) S^{X}(\varnothing) S X ( ∅ ) if U 0 U_0 U 0 is empty – and of all slices that can be reached from it via references, in an order such that each slice occurs after the slices it refers to. Then κ \kappa κ is a C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core verification of Φ \Phi Φ .
Since each slice S U ( γ ~ ) S^{U}(\widetilde{\gamma}) S U ( γ ) refers to at most c c c slices and each slice S X ( γ ~ ) S^{X}(\widetilde{\gamma}) S X ( γ ) refers to only one slice, κ \kappa κ contains at most c j c^j c j slices of the block ∀ U j \forall U_j ∀ U j for j = 0 , … , m j=0,\ldots,m j = 0 , … , m , each of size at most c c c , at most c j c^j c j slices of the block ∃ X j \exists X_j ∃ X j for j = 1 , … , m j=1,\ldots,m j = 1 , … , m , each of size 1, and at most c m + 1 c^{m+1} c m + 1 propositional slices of size 1. Hence, ∣ κ ∣ ≤ ∑ j = 0 m c j + 1 + ∑ j = 1 m c j + c m + 1 ≤ ( 2 m + 2 ) ⋅ c m + 1 |\kappa| \le\sum_{j=0}^{m} c^{j+1} + \sum_{j=1}^{m} c^j + c^{m+1} \le(2m+2)\cdot c^{m+1} ∣ κ ∣ ≤ ∑ j = 0 m c j + 1 + ∑ j = 1 m c j + c m + 1 ≤ ( 2 m + 2 ) ⋅ c m + 1 , and for c ≥ 2 c \ge2 c ≥ 2 we get ∣ κ ∣ ≤ 2 c m + 1 + c m + 1 + c m + 1 = 4 c m + 1 |\kappa| \le2c^{m+1} + c^{m+1} + c^{m+1} = 4c^{m+1} ∣ κ ∣ ≤ 2 c m + 1 + c m + 1 + c m + 1 = 4 c m + 1 . If U 0 U_0 U 0 is empty, the same count gives ∣ κ ∣ ≤ ( 2 m + 2 ) ⋅ c m |\kappa| \le(2m+2)\cdot c^m ∣ κ ∣ ≤ ( 2 m + 2 ) ⋅ c m .
From Theorem 29 and Proposition 30 we conclude:
Corollary 31. C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core is complete on false and on true QCNFs: every false QCNF has a C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutation and every true QCNF has a C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core verification. Together with Corollary 20 , C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core is a proof system for false QCNFs and for true QCNFs.
Next we show that CAQE width can be used for lower bounds as well. The idea is simple: As the CAQE width in a sense counts the number of optimal assignments throughout the formula, and since CAQE only terminates once no more assignments can be tested that are not worse than the previously tested ones, we need to test at least the optimal assignments before returning the truth value.
Proposition 32. For each C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutation κ \kappa κ of a false QCNF Φ \Phi Φ we have cw − ( Φ ) ≤ ∣ κ ∣ \operatorname{cw}^{-}(\Phi) \le|\kappa| cw − ( Φ ) ≤ ∣ κ ∣ . Analogously, for each true QCNF Φ \Phi Φ and C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core verification κ \kappa κ of Φ \Phi Φ we have cw + ( Φ ) ≤ ∣ κ ∣ \operatorname{cw}^{+}(\Phi) \le|\kappa| cw + ( Φ ) ≤ ∣ κ ∣ .
Proof. We first consider the case that Φ \Phi Φ is false. For m = 0 m=0 m = 0 we have cw − ( Φ ) = 1 ≤ ∣ κ ∣ \operatorname{cw}^{-}(\Phi)=1 \le|\kappa| cw − ( Φ ) = 1 ≤ ∣ κ ∣ , so let m ≥ 1 m \ge1 m ≥ 1 . Let s κ s_{\kappa} s κ be the universal winning strategy that we extract from κ \kappa κ in the proof of Proposition 18 . By (1), there are some j ∈ { 1 , … , m } j \in\{1,\ldots,m\} j ∈ { 1 , … , m } and an optimal chain δ = ( η 1 , … , η j − 1 ) ∈ K j − 1 ( Φ , s κ ) \delta= (\eta_1,\ldots,\eta_{j-1}) \in K_{j-1}(\Phi,s_{\kappa}) δ = ( η 1 , … , η j − 1 ) ∈ K j − 1 ( Φ , s κ ) such that
cw − ( Φ ) ≤ w ( Ψ ) for Ψ : = Φ ∣ game ( s κ , δ , U j − 1 ) . (2) \operatorname{cw}^{-}(\Phi) \le w(\Psi) \qquad\text{for} \qquad\Psi:= \Phi|_{\operatorname{game}(s_{\kappa},\delta,U_{j-1})}. \tag*{(2)} cw − ( Φ ) ≤ w ( Ψ ) for Ψ := Φ ∣ game ( s κ , δ , U j − 1 ) . (2)
If the existential player plays δ \delta δ , then the extraction of s κ s_{\kappa} s κ follows a path of slices through κ \kappa κ that leads to a slice S S S of the block ∃ X j \exists X_j ∃ X j . We show that S S S is a slice for Ψ \Psi Ψ . More generally, the slice of the block ∀ U i \forall U_i ∀ U i on this path is a slice for Φ ∣ game ( s κ , δ , X i ) \Phi|_{\operatorname{game}(s_{\kappa},\delta,X_i)} Φ ∣ game ( s κ , δ , X i ) , and the slice of the block ∃ X i \exists X_i ∃ X i is a slice for Φ ∣ game ( s κ , δ , U i − 1 ) \Phi|_{\operatorname{game}(s_{\kappa},\delta,U_{i-1})} Φ ∣ game ( s κ , δ , U i − 1 ) . This holds at the beginning, since the last slice of κ \kappa κ is a slice for Φ \Phi Φ . In a slice of a universal block with the formula Ψ ′ \Psi' Ψ ′ , the extraction plays the last assignment α g \alpha_g α g and continues with the slice for Ψ ′ ∣ β g \Psi'|_{\beta_g} Ψ ′ ∣ β g , which is Ψ ′ ∣ α g \Psi'|_{\alpha_g} Ψ ′ ∣ α g by Lemma 13. In a slice of an existential block X i X_i X i with the formula Ψ ′ = Φ ∣ game ( s κ , δ , U i − 1 ) \Psi'=\Phi|_{\operatorname{game}(s_{\kappa},\delta,U_{i-1})} Ψ ′ = Φ ∣ game ( s κ , δ , U i − 1 ) , the extraction selects some h h h with β h − 1 ( 1 ) ⊆ β γ − 1 ( 1 ) \beta_h^{-1}(1)\subseteq\beta_{\gamma}^{-1}(1) β h − 1 ( 1 ) ⊆ β γ − 1 ( 1 ) , where β γ \beta_{\gamma} β γ is canonical for η i \eta_i η i . By Lemma 17, α h \alpha_h α h satisfies all clauses of Ψ ′ \Psi' Ψ ′ that are satisfied by η i \eta_i η i . As η i \eta_i η i is an optimal move for Ψ ′ \Psi' Ψ ′ , the two assignments satisfy the same clauses, and Lemma 17 gives us Ψ ′ ∣ β h = Ψ ′ ∣ η i \Psi'|_{\beta_h}=\Psi'|_{\eta_i} Ψ ′ ∣ β h = Ψ ′ ∣ η i for the formula of the next slice.
Now let α 1 , … , α g \alpha_1,\ldots,\alpha_g α 1 , … , α g be the assignments of S S S , and let M M M be a ⊆ \subseteq ⊆ -maximal element of { sat Ψ ( α ) : α ∈ ⟨ X j ⟩ } \{\operatorname{sat}_{\Psi}(\alpha):\alpha\in\langle X_j\rangle\} { sat Ψ ( α ) : α ∈ ⟨ X j ⟩} , say M = sat Ψ ( γ ) M=\operatorname{sat}_{\Psi}(\gamma) M = sat Ψ ( γ ) . By Lemma 17, there is some h h h with M ⊆ sat Ψ ( α h ) M\subseteq\operatorname{sat}_{\Psi}(\alpha_h) M ⊆ sat Ψ ( α h ) , and we get M = sat Ψ ( α h ) M=\operatorname{sat}_{\Psi}(\alpha_h) M = sat Ψ ( α h ) by the maximality of M M M . Therefore, different ⊆ \subseteq ⊆ -maximal sets belong to different assignments of S S S , and we conclude
cw − ( Φ ) ≤ w ( Ψ ) ≤ ∣ S ∣ ≤ ∣ κ ∣ . \operatorname{cw}^{-}(\Phi)\leq w(\Psi)\leq|S|\leq|\kappa|. cw − ( Φ ) ≤ w ( Ψ ) ≤ ∣ S ∣ ≤ ∣ κ ∣.
The case for true formulas Φ \Phi Φ works analogously, with the roles of the existential and the universal blocks swapped. We use the existential winning strategy s κ s_{\kappa} s κ from Proposition 19 and obtain some j ∈ { 0 , … , m } j\in\{0,\ldots,m\} j ∈ { 0 , … , m } and an optimal chain δ ∈ K j − 1 ( Φ , s κ ) \delta\in K_{j-1}(\Phi,s_{\kappa}) δ ∈ K j − 1 ( Φ , s κ ) with cw + ( Φ ) ≤ w ( Φ ′ ) \operatorname{cw}^{+}(\Phi)\leq w(\Phi') cw + ( Φ ) ≤ w ( Φ ′ ) for Φ ′ = Φ ∣ game ( s κ , δ , X j ) \Phi'=\Phi|_{\operatorname{game}(s_{\kappa},\delta,X_j)} Φ ′ = Φ ∣ game ( s κ , δ , X j ) (if j = 0 j=0 j = 0 and U 0 U_0 U 0 is empty, the bound is trivial). If the universal player plays δ \delta δ , the extraction arrives at a verification slice S S S of the block ∀ U j \forall U_j ∀ U j . There is only one difference. By Lemma 14 and Lemma 17, the formulas of the slices on this path contain the corresponding formulas Φ ∣ game ( s κ , δ , ⋅ ) \Phi|_{\operatorname{game}(s_{\kappa},\delta,\cdot)} Φ ∣ game ( s κ , δ , ⋅ ) , but they may contain further clauses. Hence, S S S is a slice for a formula Ψ ⊇ Φ ′ \Psi\supseteq\Phi' Ψ ⊇ Φ ′ . This does not affect the bound. Let M M M be a ⊆ \subseteq ⊆ -minimal element of { sat Φ ′ ( α ) : α ∈ ⟨ U j ⟩ } \{\operatorname{sat}_{\Phi'}(\alpha):\alpha\in\langle U_j\rangle\} { sat Φ ′ ( α ) : α ∈ ⟨ U j ⟩} , say M = sat Φ ′ ( γ ) M=\operatorname{sat}_{\Phi'}(\gamma) M = sat Φ ′ ( γ ) . By Lemma 17, there is an assignment α h \alpha_h α h of S S S such that every clause of Ψ \Psi Ψ that is satisfied by α h \alpha_h α h is also satisfied by γ \gamma γ . In particular, we have sat Φ ′ ( α h ) ⊆ M \operatorname{sat}_{\Phi'}(\alpha_h)\subseteq M sat Φ ′ ( α h ) ⊆ M and hence sat Φ ′ ( α h ) = M \operatorname{sat}_{\Phi'}(\alpha_h)=M sat Φ ′ ( α h ) = M by the minimality of M M M . Again, different sets M M M belong to different assignments of S S S , which gives cw + ( Φ ) ≤ w ( Φ ′ ) ≤ ∣ S ∣ ≤ ∣ κ ∣ \operatorname{cw}^{+}(\Phi)\leq w(\Phi')\leq|S|\leq|\kappa| cw + ( Φ ) ≤ w ( Φ ′ ) ≤ ∣ S ∣ ≤ ∣ κ ∣ . ◀
We now define new QBFs Copy n \operatorname{Copy}_n Copy n consisting of n n n disjoint copies of E q 1 \mathrm{Eq}_1 Eq 1 :
Definition 33. The false QCNF Copy n \operatorname{Copy}_n Copy n has prefix ∃ X ∀ U ∃ T \exists X\forall U\exists T ∃ X ∀ U ∃ T with X = { x 1 , … , x n } X=\{x_1,\ldots,x_n\} X = { x 1 , … , x n } , U = { u 1 , … , u n } U=\{u_1,\ldots,u_n\} U = { u 1 , … , u n } and T = { t 1 , … , t n } T=\{t_1,\ldots,t_n\} T = { t 1 , … , t n } and clauses x i ∨ u i ∨ t i x_i\lor u_i\lor t_i x i ∨ u i ∨ t i , x ˉ i ∨ u ˉ i ∨ t i \bar{x}_i\lor\bar{u}_i\lor t_i x ˉ i ∨ u ˉ i ∨ t i and ( t ˉ i ) (\bar{t}_i) ( t ˉ i ) for i = 1 , … , n i=1,\ldots,n i = 1 , … , n .
These QBFs witness an exponential separation between C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core and ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res :
Theorem 34. Copy n \operatorname{Copy}_n Copy n needs C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutations of size at least 2 n 2^n 2 n , but has constant-size ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res (and ∀ R e d + R e s N P \forall\mathrm{Red}+\mathrm{Res}^{\mathrm{NP}} ∀ Red + Res NP ) refutations.
Proof. For the C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core lower bound, we can simply compute the CAQE width of Copy n \operatorname{Copy}_n Copy n . We obtain cw − ( Copy n ) = 2 n \operatorname{cw}^{-}(\operatorname{Copy}_n)=2^n cw − ( Copy n ) = 2 n by the same argument as in Proposition 26. Two different assignments of X X X differ in some x i x_i x i , so one of them satisfies x i ∨ u i ∨ t i x_i\lor u_i\lor t_i x i ∨ u i ∨ t i but not x ˉ i ∨ u ˉ i ∨ t i \bar{x}_i\lor\bar{u}_i\lor t_i x ˉ i ∨ u ˉ i ∨ t i , and vice versa for the other one. Hence, each C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutation of Copy n \operatorname{Copy}_n Copy n has at least size 2 n 2^n 2 n by Proposition 32. For the ∀ R e d + R e s \forall\mathrm{Red}+\mathrm{Res} ∀ Red + Res refutation, let C 1 = x 1 ∨ u 1 ∨ t 1 C_1=x_1\lor u_1\lor t_1 C 1 = x 1 ∨ u 1 ∨ t 1 , C 2 = x ˉ 1 ∨ u ˉ 1 ∨ t 1 C_2=\bar{x}_1\lor\bar{u}_1\lor t_1 C 2 = x ˉ 1 ∨ u ˉ 1 ∨ t 1 and C 3 = ( t ˉ 1 ) C_3=(\bar{t}_1) C 3 = ( t ˉ 1 ) . Then we can do the resolution steps
{ 1 } T { 3 } T { 1 , 3 } U π and { 2 } T { 3 } T { 2 , 3 } U π , (3) \frac{\{1\}^{T}\qquad\{3\}^{T}}{\{1,3\}^{U}}\ \pi
\qquad\text{and}\qquad
\frac{\{2\}^{T}\qquad\{3\}^{T}}{\{2,3\}^{U}}\ \pi, \tag*{(3)} { 1 , 3 } U { 1 } T { 3 } T π and { 2 , 3 } U { 2 } T { 3 } T π , (3)
where π \pi π is the Res refutation of ( t 1 ) ∧ ( t ‾ 1 ) (t_1) \wedge(\overline{t}_1) ( t 1 ) ∧ ( t 1 ) ; note that ⋁ q ∈ { 1 } C q ∣ T = ( t 1 ) \bigvee_{q \in\{1\}} C_q\vert_T = (t_1) ⋁ q ∈ { 1 } C q ∣ T = ( t 1 ) and ⋁ q ∈ { 3 } C q ∣ T = ( t ‾ 1 ) \bigvee_{q \in\{3\}} C_q\vert_T = (\overline{t}_1) ⋁ q ∈ { 3 } C q ∣ T = ( t 1 ) . The clauses C 1 , C 3 C_1,C_3 C 1 , C 3 (resp. C 2 , C 3 C_2,C_3 C 2 , C 3 ) do not contain complementary literals from U U U , and therefore we can reduce both sets
{ 1 , 3 } U { 1 , 3 } X and { 2 , 3 } U { 2 , 3 } X \frac{\{1,3\}^{U}}{\{1,3\}^{X}}
\qquad\text{and}\qquad
\frac{\{2,3\}^{U}}{\{2,3\}^{X}} { 1 , 3 } X { 1 , 3 } U and { 2 , 3 } X { 2 , 3 } U
and finish with one last resolution step
{ 1 , 3 } X { 2 , 3 } X { 1 , 2 , 3 } O π ′ , \frac{\{1,3\}^{X}\qquad\{2,3\}^{X}}{\{1,2,3\}^{O}}\ \pi', { 1 , 2 , 3 } O { 1 , 3 } X { 2 , 3 } X π ′ ,
where π ′ \pi' π ′ is the Res refutation of ( x 1 ) ∧ ( x ‾ 1 ) = ⋁ q ∈ { 1 , 3 } C q ∣ X ∧ ⋁ q ∈ { 2 , 3 } C q ∣ X (x_1)\wedge(\overline{x}_1)=\bigvee_{q\in\{1,3\}} C_q\vert_X\wedge\bigvee_{q\in\{2,3\}} C_q\vert_X ( x 1 ) ∧ ( x 1 ) = ⋁ q ∈ { 1 , 3 } C q ∣ X ∧ ⋁ q ∈ { 2 , 3 } C q ∣ X . This ∀ \forall ∀ Red + Res refutation consists of 8 lines and uses Res refutations of constant size, independently of n n n . ◀
Consequently, C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core is simulated by ∀ \forall ∀ Red + R e s N P \mathrm{Res}^{\mathrm{NP}} Res NP , but exponentially weaker. Hence the system ∀ \forall ∀ Red + Res is too strong to capture the basic version of CAQE without failed assumptions. Whether ∀ \forall ∀ Red + Res correctly captures the strength of full CAQE with failed assumptions remains open.
CAQE vs QN P ^{\mathrm{NP}} NP Res: CAQE width and unified decision lists
In this section we want to show that C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core is simulated by Q N P Q^{\mathrm{NP}} Q NP Res. While versions of Q-Res (without the NP oracle) underlie proof systems for QCDCL [16 ] , the connection between Q N P Q^{\mathrm{NP}} Q NP Res and C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core appears nontrivial. This is why we will make use of decision lists , which represent strategies and even characterise Q N P Q^{\mathrm{NP}} Q NP Res on bounded quantifier alternation [5 ] .
Definition 35 (decision lists, [36 , 5 ] ). A multi-output decision list M \mathcal{M} M from X X X to U U U , where X X X and U U U are disjoint sets of variables, is a sequence of pairs ( D 1 , μ 1 ) , … , ( D z , μ z ) (D_1,\mu_1),\ldots,(D_z,\mu_z) ( D 1 , μ 1 ) , … , ( D z , μ z ) where
D i D_i D i are cubes with var ( D i ) ⊆ X \operatorname{var}(D_i)\subseteq X var ( D i ) ⊆ X and ⋁ i = 1 z D i ≡ ⊤ \bigvee_{i=1}^{z}D_i\equiv\top ⋁ i = 1 z D i ≡ ⊤ ,
μ i ∈ ⟨ U ⟩ \mu_i\in\langle U\rangle μ i ∈ ⟨ U ⟩
for i = 1 , … , z i=1,\ldots,z i = 1 , … , z . The size of M \mathcal{M} M is defined as ∣ M ∣ : = z |\mathcal{M}|:=z ∣ M ∣ := z .
M \mathcal{M} M computes the function from ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ to ⟨ U ⟩ \langle U\rangle ⟨ U ⟩ that maps each γ ∈ ⟨ X ⟩ \gamma\in\langle X\rangle γ ∈ ⟨ X ⟩ to μ i γ \mu_{i_\gamma} μ i γ , where i γ : = min { i : γ ⊨ D i } i_\gamma:=\min\{i:\gamma\models D_i\} i γ := min { i : γ ⊨ D i } . If M \mathcal{M} M computes a universal winning strategy for a QCNF Φ \Phi Φ , then M \mathcal{M} M is called a unified decision list (UDL) for Φ \Phi Φ . That means the function γ ↦ M ( γ ) ∣ U j \gamma\mapsto\mathcal{M}(\gamma)\vert_{U_j} γ ↦ M ( γ ) ∣ U j only depends on γ ∣ X 1 ∪ … ∪ X j \gamma\vert_{X_1\cup\ldots\cup X_j} γ ∣ X 1 ∪ … ∪ X j for each j j j , and the resulting strategy is winning. In that case, we call M \mathcal{M} M a UDL refutation of Φ \Phi Φ . We will use the name of the UDL M \mathcal{M} M to refer to the computed function.
Theorem 36 ([5 Theorem 4.2 and Corollary 5.6] ). Q N P Q^{\mathrm{NP}} Q NP Res ≡ p \equiv_p ≡ p UDL on bounded quantifier alternation.
Instead of proving that Q N P Q^{\mathrm{NP}} Q NP Res simulates C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core directly, we show that UDLs are bounded by the CAQE width.
Theorem 37. Let Φ \Phi Φ be a false QCNF. Then there exists a UDL M M M for Φ \Phi Φ with ∣ M ∣ ≤ cw − ( Φ ) m |M|\leq\operatorname{cw}^{-}(\Phi)^m ∣ M ∣ ≤ cw − ( Φ ) m .
Proof. Let s s s be a universal winning strategy for Φ \Phi Φ that attains the minimum in Definition 23. For j ≤ m j\leq m j ≤ m and an assignment σ \sigma σ of X 1 ∪ … ∪ X j X_1\cup\ldots\cup X_j X 1 ∪ … ∪ X j (or of a superset of these variables) we abbreviate sat j ( σ ) : = { r : game ( s , σ , X j ) ⊨ C r } \operatorname{sat}_j(\sigma):=\{r:\operatorname{game}(s,\sigma,X_j)\models C_r\} sat j ( σ ) := { r : game ( s , σ , X j ) ⊨ C r } , which is the set of clauses satisfied in the game up to block X j X_j X j .
We fix an arbitrary linear order on each ⟨ X j ⟩ \langle X_j\rangle ⟨ X j ⟩ , and lexicographically minimal moves and chains will always refer to these orders.
Let σ = ( σ 1 , … , σ m + 1 ) ∈ ⟨ X ⟩ \sigma=(\sigma_1,\ldots,\sigma_{m+1})\in\langle X\rangle σ = ( σ 1 , … , σ m + 1 ) ∈ ⟨ X ⟩ be arbitrary. We define an optimal chain η ( σ ) : = ( η 1 , … , η m ) ∈ K ( Φ , s ) \eta(\sigma):=(\eta_1,\ldots,\eta_m)\in K(\Phi,s) η ( σ ) := ( η 1 , … , η m ) ∈ K ( Φ , s ) inductively: Let η 1 ∈ ⟨ X 1 ⟩ \eta_1\in\langle X_1\rangle η 1 ∈ ⟨ X 1 ⟩ be the lexicographically minimal optimal move for Φ ∣ game ( s , ∅ , U 0 ) \Phi|_{\operatorname{game}(s,\emptyset,U_0)} Φ ∣ game ( s , ∅ , U 0 ) such that sat 1 ( σ 1 ) ⊆ sat 1 ( η 1 ) \operatorname{sat}_1(\sigma_1)\subseteq\operatorname{sat}_1(\eta_1) sat 1 ( σ 1 ) ⊆ sat 1 ( η 1 ) . This is possible as either σ 1 \sigma_1 σ 1 is already an optimal move, or there exists an optimal move η 1 \eta_1 η 1 that satisfies all clauses satisfied by σ 1 \sigma_1 σ 1 (cf. the discussion after Definition 21). Assume that we already constructed η 1 , … , η j \eta_1,\ldots,\eta_j η 1 , … , η j . Then let η j + 1 ∈ ⟨ X j + 1 ⟩ \eta_{j+1}\in\langle X_{j+1}\rangle η j + 1 ∈ ⟨ X j + 1 ⟩ be the lexicographically minimal optimal move for Φ ∣ game ( s , η 1 ∪ … ∪ η j , U j ) \Phi|_{\operatorname{game}(s,\eta_1\cup\ldots\cup\eta_j,U_j)} Φ ∣ game ( s , η 1 ∪ … ∪ η j , U j ) such that sat j + 1 ( η 1 ∪ … ∪ η j ∪ σ j + 1 ) ⊆ sat j + 1 ( η 1 ∪ … ∪ η j + 1 ) \operatorname{sat}_{j+1}(\eta_1\cup\ldots\cup\eta_j\cup\sigma_{j+1})\subseteq\operatorname{sat}_{j+1}(\eta_1\cup\ldots\cup\eta_{j+1}) sat j + 1 ( η 1 ∪ … ∪ η j ∪ σ j + 1 ) ⊆ sat j + 1 ( η 1 ∪ … ∪ η j + 1 ) . Such an η j + 1 \eta_{j+1} η j + 1 exists by the same argument. Note that η 1 , … , η j \eta_1,\ldots,\eta_j η 1 , … , η j only depend on σ 1 , … , σ j \sigma_1,\ldots,\sigma_j σ 1 , … , σ j by construction. We have now constructed η ( σ ) = ( η 1 , … , η m ) ∈ K ( Φ , s ) \eta(\sigma)=(\eta_1,\ldots,\eta_m)\in K(\Phi,s) η ( σ ) = ( η 1 , … , η m ) ∈ K ( Φ , s ) .
We define a strategy s ~ = ( s ~ U 0 , … , s ~ U m ) \tilde{s}=(\tilde{s}_{U_0},\ldots,\tilde{s}_{U_m}) s ~ = ( s ~ U 0 , … , s ~ U m ) by s ~ U j ( σ 1 ∪ … ∪ σ j ) : = s U j ( η 1 ∪ … ∪ η j ) \tilde{s}_{U_j}(\sigma_1\cup\ldots\cup\sigma_j):=s_{U_j}(\eta_1\cup\ldots\cup\eta_j) s ~ U j ( σ 1 ∪ … ∪ σ j ) := s U j ( η 1 ∪ … ∪ η j ) , where ( η 1 , … , η j ) (\eta_1,\ldots,\eta_j) ( η 1 , … , η j ) is the prefix of η ( σ ) \eta(\sigma) η ( σ ) . As we noted above, η 1 , … , η j \eta_1,\ldots,\eta_j η 1 , … , η j only depend on σ 1 , … , σ j \sigma_1,\ldots,\sigma_j σ 1 , … , σ j , so this is indeed a strategy. It remains to show that it is in fact a winning strategy, meaning that for all σ ∈ ⟨ X ⟩ \sigma\in\langle X\rangle σ ∈ ⟨ X ⟩ the assignment σ ∪ s ~ ( σ ) \sigma\cup\tilde{s}(\sigma) σ ∪ s ~ ( σ ) has to falsify some clause C q C_q C q .
Let σ = ( σ 1 , … , σ m + 1 ) ∈ ⟨ X ⟩ \sigma=(\sigma_1,\ldots,\sigma_{m+1})\in\langle X\rangle σ = ( σ 1 , … , σ m + 1 ) ∈ ⟨ X ⟩ be arbitrary, let η ( σ ) = ( η 1 , … , η m ) \eta(\sigma)=(\eta_1,\ldots,\eta_m) η ( σ ) = ( η 1 , … , η m ) and consider the assignment γ : = ( η 1 , … , η m , σ m + 1 ) ∈ ⟨ X ⟩ \gamma:=(\eta_1,\ldots,\eta_m,\sigma_{m+1})\in\langle X\rangle γ := ( η 1 , … , η m , σ m + 1 ) ∈ ⟨ X ⟩ . The value s ( γ ) s(\gamma) s ( γ ) only depends on η 1 , … , η m \eta_1,\ldots,\eta_m η 1 , … , η m , so we have s ( γ ) = s ~ ( σ ) s(\gamma)=\tilde{s}(\sigma) s ( γ ) = s ~ ( σ ) . We know that s s s is a winning strategy, hence some C q C_q C q is falsified by γ ∪ s ( γ ) \gamma\cup s(\gamma) γ ∪ s ( γ ) . Assume, for the sake of contradiction, that C q C_q C q is satisfied by σ ∪ s ~ ( σ ) = σ ∪ s ( γ ) \sigma\cup\tilde{s}(\sigma)=\sigma\cup s(\gamma) σ ∪ s ~ ( σ ) = σ ∪ s ( γ ) . Since γ ∪ s ( γ ) \gamma\cup s(\gamma) γ ∪ s ( γ ) and σ ∪ s ( γ ) \sigma\cup s(\gamma) σ ∪ s ( γ ) agree on U U U and on X m + 1 X_{m+1} X m + 1 , there is some j ≤ m j\leq m j ≤ m such that σ j ⊨ C q \sigma_j\models C_q σ j ⊨ C q . But then we have q ∈ sat j ( η 1 ∪ … ∪ η j − 1 ∪ σ j ) ⊆ sat j ( η 1 ∪ … ∪ η j ) q\in\operatorname{sat}_j(\eta_1\cup\ldots\cup\eta_{j-1}\cup\sigma_j)\subseteq\operatorname{sat}_j(\eta_1\cup\ldots\cup\eta_j) q ∈ sat j ( η 1 ∪ … ∪ η j − 1 ∪ σ j ) ⊆ sat j ( η 1 ∪ … ∪ η j ) , and therefore game ( s , η 1 ∪ … ∪ η j , X j ) ⊆ γ ∪ s ( γ ) \operatorname{game}(s,\eta_1\cup\ldots\cup\eta_j,X_j)\subseteq\gamma\cup s(\gamma) game ( s , η 1 ∪ … ∪ η j , X j ) ⊆ γ ∪ s ( γ ) would also satisfy C q C_q C q . This is a contradiction to the fact that C q C_q C q is falsified by γ ∪ s ( γ ) \gamma\cup s(\gamma) γ ∪ s ( γ ) . Hence s ~ \tilde{s} s ~ is indeed a winning strategy for Φ \Phi Φ . Intuitively, s ~ \tilde{s} s ~ projects all input assignments to better optimal, lexicographically minimal assignments and applies s s s to these optimal assignments.
Let H : = { η ( σ ) : σ ∈ ⟨ X ⟩ } ⊆ K ( Φ , s ) H:=\{\eta(\sigma):\sigma\in\langle X\rangle\}\subseteq K(\Phi,s) H := { η ( σ ) : σ ∈ ⟨ X ⟩} ⊆ K ( Φ , s ) , and let γ 1 < γ 2 < … < γ z \gamma_1<\gamma_2<\ldots<\gamma_z γ 1 < γ 2 < … < γ z be the elements of H H H in lexicographic order. We first show z ≤ cw − ( Φ ) m z\leq\operatorname{cw}^{-}(\Phi)^m z ≤ cw − ( Φ ) m . For j ≤ m j\leq m j ≤ m let H j H_j H j be the set of the prefixes ( η 1 , … , η j ) (\eta_1,\ldots,\eta_j) ( η 1 , … , η j ) of the chains in H H H . We have ∣ H 0 ∣ = 1 |H_0|=1 ∣ H 0 ∣ = 1 , so it suffices to show that every θ = ( η 1 , … , η j − 1 ) ∈ H j − 1 \theta=(\eta_1,\ldots,\eta_{j-1})\in H_{j-1} θ = ( η 1 , … , η j − 1 ) ∈ H j − 1 has at most cw − ( Φ ) \operatorname{cw}^{-}(\Phi) cw − ( Φ ) extensions ( θ , η j ) ∈ H j (\theta,\eta_j)\in H_j ( θ , η j ) ∈ H j . By the definition of η ( σ ) \eta(\sigma) η ( σ ) , each such η j \eta_j η j is the lexicographically minimal optimal move for Φ ∣ game ( s , θ , U j − 1 ) \Phi|_{\operatorname{game}(s,\theta,U_{j-1})} Φ ∣ game ( s , θ , U j − 1 ) with sat j ( θ ∪ η j ) ⊆ sat j ( θ ∪ η j ) \operatorname{sat}_j(\theta\cup\eta_j)\subseteq\operatorname{sat}_j(\theta\cup\eta_j) sat j ( θ ∪ η j ) ⊆ sat j ( θ ∪ η j ) , for some σ j ∈ ⟨ X j ⟩ \sigma_j\in\langle X_j\rangle σ j ∈ ⟨ X j ⟩ . Let ( θ , η j ) (\theta,\eta_j) ( θ , η j ) and ( θ , η j ′ ) (\theta,\eta'_j) ( θ , η j ′ ) be two such extensions with sat j ( θ ∪ η j ) ≠ sat j ( θ ∪ η j ′ ) \operatorname{sat}_j(\theta\cup\eta_j)\neq\operatorname{sat}_j(\theta\cup\eta'_j) sat j ( θ ∪ η j ) = sat j ( θ ∪ η j ′ ) . Then η j ′ \eta'_j η j ′ is one of the optimal moves among which η j \eta_j η j was chosen, so η j ≤ η j ′ \eta_j\leq\eta'_j η j ≤ η j ′ , and in the same way η j ′ ≤ η j \eta'_j\leq\eta_j η j ′ ≤ η j . Hence, different extensions have different sets sat j ( θ ∪ η j ) \operatorname{sat}_j(\theta\cup\eta_j) sat j ( θ ∪ η j ) . Since the η j \eta_j η j are optimal moves, these sets are ⊆ \subseteq ⊆ -maximal elements of { sat j ( γ ) : θ ⊆ γ ∈ ⟨ X ⟩ } \{\operatorname{sat}_j(\gamma):\theta\subseteq\gamma\in\langle X\rangle\} { sat j ( γ ) : θ ⊆ γ ∈ ⟨ X ⟩} . As we have θ ∈ K j − 1 ( Φ , s ) \theta\in K_{j-1}(\Phi,s) θ ∈ K j − 1 ( Φ , s ) and s s s attains the minimum in Definition 23, there are at most cw − ( Φ ) \operatorname{cw}^{-}(\Phi) cw − ( Φ ) of them.
For i = 1 , … , z i=1,\ldots,z i = 1 , … , z let ξ i : = fru ( γ i ) \xi_i:=\operatorname{fru}(\gamma_i) ξ i := fru ( γ i ) . We now define M \mathcal{M} M as ( D 1 , μ 1 ) , … , ( D z , μ z ) (D_1,\mu_1),\ldots,(D_z,\mu_z) ( D 1 , μ 1 ) , … , ( D z , μ z ) with
D i : = ⋀ j = 1 m ⋀ r ∉ I j ¬ ( C r ∣ X 1 ∪ … ∪ X j ) and μ i : = s ( γ i ) D_i:=\bigwedge_{j=1}^{m}\bigwedge_{r\notin I_j}\neg\left(C_r|_{X_1\cup\ldots\cup X_j}\right)
\qquad\text{and}\qquad
\mu_i:=s(\gamma_i) D i := j = 1 ⋀ m r ∈ / I j ⋀ ¬ ( C r ∣ X 1 ∪ … ∪ X j ) and μ i := s ( γ i )
for i = 1 , … , z i=1,\ldots,z i = 1 , … , z , where ( I 1 , … , I m ) = ξ i (I_1,\ldots,I_m)=\xi_i ( I 1 , … , I m ) = ξ i . In other words, for each j j j we go through all indices r r r that do not appear in the j t h j^{\mathrm{th}} j th set I j I_j I j of ξ i \xi_i ξ i . Here ¬ ( C ∣ V ) \neg(C|_V) ¬ ( C ∣ V ) denotes the cube consisting of the negations of all literals of C ∣ V C|_V C ∣ V , which is the empty cube if C ∣ V C|_V C ∣ V is empty. So D i D_i D i is a cube over X 1 ∪ … ∪ X m X_1\cup\ldots\cup X_m X 1 ∪ … ∪ X m , and an assignment σ ∈ ⟨ X ⟩ \sigma\in\langle X\rangle σ ∈ ⟨ X ⟩ satisfies D i D_i D i if and only if σ ∪ … ∪ σ j \sigma\cup\ldots\cup\sigma_j σ ∪ … ∪ σ j satisfies no clause C r C_r C r with r ∉ I j r\notin I_j r ∈ / I j , for all j = 1 , … , m j=1,\ldots,m j = 1 , … , m . It remains to show that M \mathcal{M} M is a UDL, which we will do by proving that M \mathcal{M} M computes s ~ \tilde{s} s ~ .
Claim 38. For each σ ∈ ⟨ X ⟩ \sigma\in\langle X\rangle σ ∈ ⟨ X ⟩ and i ∈ { 1 , … , z } i\in\{1,\ldots,z\} i ∈ { 1 , … , z } the following are equivalent:
(i) i i i is minimal such that σ ⊨ D i \sigma\models D_i σ ⊨ D i ;
(ii) γ i = η ( σ ) \gamma_i = \eta(\sigma) γ i = η ( σ ) .
Proof of the claim. Let σ ∈ ⟨ X ⟩ \sigma\in\langle X\rangle σ ∈ ⟨ X ⟩ be fixed. We first show the following two facts for an arbitrary i i i with ξ i = ( I 1 , … , I m ) \xi_i = (I_1,\ldots,I_m) ξ i = ( I 1 , … , I m ) and γ i = ( η 1 ′ , … , η m ′ ) \gamma_i = (\eta'_1,\ldots,\eta'_m) γ i = ( η 1 ′ , … , η m ′ ) .
Fact A: If σ ⊨ D i \sigma\models D_i σ ⊨ D i , then sat j ( η 1 ′ ∪ … ∪ η j − 1 ′ ∪ σ j ) ⊆ sat j ( η 1 ′ ∪ … ∪ η j ′ ) = I j \operatorname{sat}_j(\eta'_1 \cup\ldots\cup\eta'_{j-1} \cup\sigma_j) \subseteq\operatorname{sat}_j(\eta'_1 \cup\ldots\cup\eta'_j) = I_j sat j ( η 1 ′ ∪ … ∪ η j − 1 ′ ∪ σ j ) ⊆ sat j ( η 1 ′ ∪ … ∪ η j ′ ) = I j for all j = 1 , … , m j = 1,\ldots,m j = 1 , … , m . Let q ∈ sat j ( η 1 ′ ∪ … ∪ η j − 1 ′ ∪ σ j ) q \in\operatorname{sat}_j(\eta'_1 \cup\ldots\cup\eta'_{j-1} \cup\sigma_j) q ∈ sat j ( η 1 ′ ∪ … ∪ η j − 1 ′ ∪ σ j ) . If C q C_q C q is satisfied by the universal part of game ( s , η 1 ′ ∪ … ∪ η j − 1 ′ ∪ σ j , X j ) \operatorname{game}(s,\eta'_1 \cup\ldots\cup\eta'_{j-1} \cup\sigma_j,X_j) game ( s , η 1 ′ ∪ … ∪ η j − 1 ′ ∪ σ j , X j ) or by η 1 ′ ∪ … ∪ η j − 1 ′ \eta'_1 \cup\ldots\cup\eta'_{j-1} η 1 ′ ∪ … ∪ η j − 1 ′ , then C q C_q C q is also satisfied by game ( s , η 1 ′ ∪ … ∪ η j ′ , X j ) \operatorname{game}(s,\eta'_1 \cup\ldots\cup\eta'_j,X_j) game ( s , η 1 ′ ∪ … ∪ η j ′ , X j ) — the two assignments agree outside of X j X_j X j . In the other case, σ j \sigma_j σ j satisfies C q C_q C q , hence σ 1 ∪ … ∪ σ j \sigma_1 \cup\ldots\cup\sigma_j σ 1 ∪ … ∪ σ j satisfies C q C_q C q as well. Since σ ⊨ D i \sigma\models D_i σ ⊨ D i , we have q ∈ I j q \in I_j q ∈ I j . In both cases, we get q ∈ I j = sat j ( η 1 ′ ∪ … ∪ η j ′ ) q \in I_j = \operatorname{sat}_j(\eta'_1 \cup\ldots\cup\eta'_j) q ∈ I j = sat j ( η 1 ′ ∪ … ∪ η j ′ ) .
Fact B: If γ i = η ( σ ) \gamma_i = \eta(\sigma) γ i = η ( σ ) , then σ ⊨ D i \sigma\models D_i σ ⊨ D i . Let j ∈ { 1 , … , m } j \in\{1,\ldots,m\} j ∈ { 1 , … , m } and q ∉ I j q \notin I_j q ∈ / I j be arbitrary. We want to show that σ 1 ∪ … ∪ σ j \sigma_1 \cup\ldots\cup\sigma_j σ 1 ∪ … ∪ σ j does not satisfy C q C_q C q . Assume, for the sake of contradiction, that it does. Then there exists a minimal a ≤ j a \leq j a ≤ j such that σ a \sigma_a σ a satisfies C q C_q C q . By the definition of η ( σ ) = γ i \eta(\sigma) = \gamma_i η ( σ ) = γ i , this implies q ∈ sat a ( η 1 ′ ∪ … ∪ η a − 1 ′ ∪ σ a ) ⊆ sat a ( η 1 ′ ∪ … ∪ η a ′ ) = I a q \in\operatorname{sat}_a(\eta'_1 \cup\ldots\cup\eta'_{a-1} \cup\sigma_a) \subseteq\operatorname{sat}_a(\eta'_1 \cup\ldots\cup\eta'_a) = I_a q ∈ sat a ( η 1 ′ ∪ … ∪ η a − 1 ′ ∪ σ a ) ⊆ sat a ( η 1 ′ ∪ … ∪ η a ′ ) = I a . But I a ⊆ I j I_a \subseteq I_j I a ⊆ I j , which contradicts the choice of q q q . We conclude that σ \sigma σ falsifies C q ∣ X 1 ∪ … ∪ X j C_q|_{X_1\cup\ldots\cup X_j} C q ∣ X 1 ∪ … ∪ X j for all q ∉ I j q \notin I_j q ∈ / I j and all j j j , which means that σ ⊭ D i \sigma\not\models D_i σ ⊨ D i .
Now let i ∗ i^* i ∗ be the index with γ i ∗ = η ( σ ) \gamma_{i^*} = \eta(\sigma) γ i ∗ = η ( σ ) , which exists because η ( σ ) ∈ H \eta(\sigma) \in H η ( σ ) ∈ H . By Fact B, we have σ ⊨ D i ∗ \sigma\models D_{i^*} σ ⊨ D i ∗ . Let i ≤ i ∗ i \leq i^* i ≤ i ∗ be arbitrary with σ ⊨ D i \sigma\models D_i σ ⊨ D i and γ i = ( η 1 ′ , … , η m ′ ) \gamma_i = (\eta'_1,\ldots,\eta'_m) γ i = ( η 1 ′ , … , η m ′ ) . From γ i ∈ K ( Φ , s ) \gamma_i \in K(\Phi,s) γ i ∈ K ( Φ , s ) we know that each η j ′ \eta'_j η j ′ is an optimal move for Φ ∣ game ( s , η 1 ′ ∪ … ∪ η j − 1 ′ , U j − 1 ) \Phi|_{\operatorname{game}(s,\eta'_1\cup\ldots\cup\eta'_{j-1},U_{j-1})} Φ ∣ game ( s , η 1 ′ ∪ … ∪ η j − 1 ′ , U j − 1 ) , and Fact A tells us that sat j ( η 1 ′ ∪ … ∪ η j − 1 ′ ∪ σ j ) ⊆ sat j ( η 1 ′ ∪ … ∪ η j ′ ) \operatorname{sat}_j(\eta'_1 \cup\ldots\cup\eta'_{j-1} \cup\sigma_j) \subseteq\operatorname{sat}_j(\eta'_1 \cup\ldots\cup\eta'_j) sat j ( η 1 ′ ∪ … ∪ η j − 1 ′ ∪ σ j ) ⊆ sat j ( η 1 ′ ∪ … ∪ η j ′ ) for each j j j . Let us compare γ i \gamma_i γ i with η ( σ ) \eta(\sigma) η ( σ ) block by block. If they do not differ in any block, then γ i = η ( σ ) \gamma_i = \eta(\sigma) γ i = η ( σ ) and i = i ∗ i = i^* i = i ∗ . In the other case, let j j j be the first block where they differ. Then η 1 ′ = η 1 , … , η j − 1 ′ = η j − 1 \eta'_1 = \eta_1,\ldots,\eta'_{j-1} = \eta_{j-1} η 1 ′ = η 1 , … , η j − 1 ′ = η j − 1 , and by the definition of η j \eta_j η j as the lexicographically minimal optimal move with the above property we obtain η j < η j ′ \eta_j < \eta'_j η j < η j ′ . But then η ( σ ) < γ i \eta(\sigma) < \gamma_i η ( σ ) < γ i and i ∗ < i i^* < i i ∗ < i , which is a contradiction to i ≤ i ∗ i \leq i^* i ≤ i ∗ . Therefore, i ∗ i^* i ∗ is the minimal index with σ ⊨ D i ∗ \sigma\models D_{i^*} σ ⊨ D i ∗ , which proves the equivalence of (i) and (ii). ◁
By the claim, for each σ ∈ ⟨ X ⟩ \sigma\in\langle X\rangle σ ∈ ⟨ X ⟩ there is a minimal i i i with σ ⊨ D i \sigma\models D_i σ ⊨ D i , hence ⋁ i = 1 z D i ≡ ⊤ \bigvee_{i=1}^{z} D_i \equiv\top ⋁ i = 1 z D i ≡ ⊤ . The claim also gives us M ( σ ) = μ i σ = s ( γ i σ ) = s ( η ( σ ) ) = s ~ ( σ ) \mathcal{M}(\sigma) = \mu_{i_\sigma} = s(\gamma_{i_\sigma}) = s(\eta(\sigma)) = \widetilde{s}(\sigma) M ( σ ) = μ i σ = s ( γ i σ ) = s ( η ( σ )) = s ( σ ) , so M \mathcal{M} M computes exactly s ~ \widetilde{s} s . We already know that s ~ \widetilde{s} s is a universal winning strategy, therefore M \mathcal{M} M is a UDL with ∣ M ∣ = z ≤ cw − ( Φ ) m |\mathcal{M}| = z \leq\operatorname{cw}^{-}(\Phi)^m ∣ M ∣ = z ≤ cw − ( Φ ) m . ◀
Because of the lower bound of C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core proofs via CAQE width (Proposition 32 ) and the UDL characterisation of Q N P R e s \mathrm{Q}^{\mathrm{NP}}\mathrm{Res} Q NP Res (Theorem 36 ), as well as the fact that the QBFs C o p y n \mathrm{Copy}_n Copy n are obviously easy for Q N P R e s \mathrm{Q}^{\mathrm{NP}}\mathrm{Res} Q NP Res but hard for C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core (cf. Theorem 34 ) we conclude:
Corollary 39. Q N P R e s \mathrm{Q}^{\mathrm{NP}}\mathrm{Res} Q NP Res simulates C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core on bounded quantifier alternation, but C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core is exponentially weaker than Q N P R e s \mathrm{Q}^{\mathrm{NP}}\mathrm{Res} Q NP Res as witnessed by C o p y n \mathrm{Copy}_n Copy n .
Proof. Let κ \kappa κ be a C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutation of a false QCNF Φ \Phi Φ with m m m universal blocks (after a possibly empty block U 0 U_0 U 0 ). Using Proposition 32 and Theorem 37 , we know that Φ \Phi Φ has a UDL of size at most cw − ( Φ ) m ≤ ∣ κ ∣ m \operatorname{cw}^{-}(\Phi)^m \leq|\kappa|^m cw − ( Φ ) m ≤ ∣ κ ∣ m , which is polynomial in ∣ κ ∣ |\kappa| ∣ κ ∣ for bounded m m m . Now Theorem 36 gives us that Φ \Phi Φ has a Q N P R e s \mathrm{Q}^{\mathrm{NP}}\mathrm{Res} Q NP Res refutation of size polynomial in ∣ κ ∣ |\kappa| ∣ κ ∣ .
We will now show the separation. By Theorem 34 , C o p y n \mathrm{Copy}_n Copy n requires C A Q E C o r e \mathrm{CAQE}^{\mathrm{Core}} CAQE Core refutations of size 2 n 2^n 2 n . On the other hand, M = ( [ x 1 ] , u 1 ∧ u 2 ‾ ∧ … ∧ u n ‾ ) , ( [ x 1 ‾ ] , u 1 ‾ ∧ u 2 ‾ ∧ … ∧ u n ‾ ) \mathcal{M} = ([x_1],u_1 \land\overline{u_2} \land\ldots\land\overline{u_n}),([\overline{x_1}],\overline{u_1} \land\overline{u_2} \land\ldots\land\overline{u_n}) M = ([ x 1 ] , u 1 ∧ u 2 ∧ … ∧ u n ) , ([ x 1 ] , u 1 ∧ u 2 ∧ … ∧ u n ) is a UDL of size 2 for C o p y n \mathrm{Copy}_n Copy n . It computes the strategy u 1 ↦ x 1 u_1 \mapsto x_1 u 1 ↦ x 1 and u i ↦ 0 u_i \mapsto0 u i ↦ 0 for i ≥ 2 i \geq2 i ≥ 2 , under which the clauses x 1 ∨ u 1 ∨ t 1 x_1 \lor u_1 \lor t_1 x 1 ∨ u 1 ∨ t 1 , x 1 ∨ u 1 ∨ t 1 x_1 \lor u_1 \lor t_1 x 1 ∨ u 1 ∨ t 1 and ( t 1 ‾ ) (\overline{t_1}) ( t 1 ) cannot be satisfied simultaneously. Hence, C o p y n \mathrm{Copy}_n Copy n has Q N P R e s \mathrm{Q}^{\mathrm{NP}}\mathrm{Res} Q NP Res refutations of size polynomial in n n n . In fact, we can directly obtain a Q-Res refutation of constant size from the V R e d + R e s \mathrm{VRed}+\mathrm{Res} VRed + Res refutation in Theorem 34 . ◀
We remark that the simulation in Corollary 39 is a simulation in terms of proof size. The UDL in Theorem 37 is constructed from an optimal strategy s s s and from the lexicographically minimal optimal chains, and we do not claim that it can be computed from κ \kappa κ in polynomial time.
CAQE vs QCDCL
Finally, we compare C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE and QCDCL. We will show that there is a separation between C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE and QCDCL in one direction. For the other direction, at least on a specific but relevant class of QCNFs, namely Σ 3 b \Sigma^{b}_{3} Σ 3 b formulas, QCDCL can simulate C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE proofs (after a simple preprocessing step). We utilise the QCDCL proof system framework used in [19 , 16 , 8 ] that proof-theoretically tightly captures the algorithmic QCDCL paradigm. Intuitively, a QCDCL proof is a sequence of tuples of the form ( T i , C i , π i ) (T_i,C_i,\pi_i) ( T i , C i , π i ) . The T i T_i T i are called trails , which are (ordered) assignments containing all decided and propagated literals in the order in which they were assigned. Each C i C_i C i is the clause learned from the trail T i T_i T i and π i \pi_i π i is the derivation of the learned clause C i C_i C i from the axioms and previously learned clauses (e.g. via Q-Res). Below we recall the precise definitions from [8 ] .
We have already shown that the equality formulas have exponential CAQE width, hence they need exponential-size C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE proofs by Proposition 32 . We recall that E q n \mathrm{Eq}_n Eq n have QCDCL refutations of quadratic size if cube learning is used [18 Proposition 5.2] – the construction given there uses O ( n ) O(n) O ( n ) trails of length O ( n ) O(n) O ( n ) . These refutations learn cubes from total assignments; without cube learning, E q n \mathrm{Eq}_n Eq n requires exponential-size QCDCL refutations [7 ] . The QCDCL proof system that we recall below (Definition 46 ) includes cube learning. Hence we obtain:
Theorem 40. E q n \mathrm{Eq}_n Eq n require C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE refutations of size at least 2 n 2^n 2 n and have QCDCL refutations (with cube learning) of size O ( n 2 ) O(n^2) O ( n 2 ) . Hence, C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE does not simulate QCDCL.
Proof. Due to Proposition 26 we have cw − ( E q n ) = 2 n \operatorname{cw}^{-}(\mathrm{Eq}_n)=2^n cw − ( Eq n ) = 2 n . Hence, each C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE refutation of E q n \mathrm{Eq}_n Eq n has size at least 2 n 2^n 2 n by Proposition 32 . The QCDCL refutations of size O ( n 2 ) O(n^2) O ( n 2 ) follow from the construction in the proof of [18 Proposition 5.2] , as explained above.
Before we turn to the other direction, let us briefly recall how QCDCL works [41 , 14 ] . Like CDCL, the algorithm builds up a partial assignment – the trail – by alternately deciding variables and performing unit propagation. There are two differences to the propositional case. First, decisions have to respect the prefix: only variables from the leftmost quantifier block that is not yet completely assigned may be decided. Second, unit propagation is combined with universal reduction. A clause is considered unit if all its literals except one existential literal are falsified, where universal literals quantified right of this existential literal are ignored. In particular, universal literals are never propagated from clauses. If the trail falsifies a clause, we have run into a conflict . As in CDCL, the algorithm then analyses the conflict: it resolves the falsified clause with the antecedent clauses of propagated literals in reverse trail order, learns the resulting clause (after universal reduction) and backtracks to an earlier point of the trail. In this way, the learned clauses form a derivation in long-distance Q-resolution, and the QCNF is refuted as soon as the empty clause is learned. Symmetrically, QCDCL can learn cubes from trails that satisfy the matrix, which is needed to verify true QCNFs. Cube learning is only needed for the separation in Theorem 40 ; all refutations that we construct ourselves in this section learn clauses only.
One major difference between C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE and QCDCL is that QCDCL does not directly call a SAT solver. Instead, QCDCL itself is a natural extension of CDCL and therefore contains all CDCL functionality. This also means that it is not as trivial to discard propositional hardness from QCDCL as it is with C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE where computations by the SAT oracle are not counted towards C A Q E C O R E \mathrm{CAQE}^{\mathrm{CORE}} CAQE CORE proof size. Hence we need to define explicitly which steps of QCDCL could be performed by CDCL as well. Informally, a CDCL step in a QCDCL proof ι \iota ι is a trail of ι \iota ι (together with its learned clause) that can occur in a CDCL proof as well. Such a CDCL step occurs if the formula under the current assignment (i.e. the trail) contains a subformula that is propositional, unsatisfiable and only contains variables from the (currently) first existential quantifier block. In such a situation, QCDCL can run into a conflict by propagating and deciding variables from this subformula only without any reductions. The subsequent clause learning will therefore only perform operations that are also doable in CDCL if we do not overstep the point at which the unsatisfiable subformula appeared. In order to formalise this, we first need to recall the QCDCL proof system from [8 ] .
Definition 41 (quantification level). Let Φ : = Q 1 V 1 Q 2 V 2 … Q d V d ⋅ φ \Phi:= Q_{1}V_{1}Q_{2}V_{2}\ldots Q_{d}V_{d}\cdot\varphi Φ := Q 1 V 1 Q 2 V 2 … Q d V d ⋅ φ be a QBF with the quantifier blocks V 1 , … , V d V_{1},\ldots,V_{d} V 1 , … , V d (pairwise disjoint sets of variables) and quantifiers Q i ∈ { ∃ , ∀ } Q_{i}\in\{\exists,\forall\} Q i ∈ { ∃ , ∀ } such that Q j ≠ Q j + 1 Q_{j}\ne Q_{j+1} Q j = Q j + 1 for all j j j . For some variable v v v from Φ \Phi Φ we set lv ( v ) : = i \operatorname{lv}(v):=i lv ( v ) := i if v ∈ V i v\in V_{i} v ∈ V i . Furthermore, we set lv ( v ‾ ) : = lv ( v ) \operatorname{lv}(\overline{v}):=\operatorname{lv}(v) lv ( v ) := lv ( v ) . For literals ℓ 1 , ℓ 2 \ell_{1},\ell_{2} ℓ 1 , ℓ 2 we write ℓ 1 < Φ ℓ 2 \ell_{1}<_{\Phi}\ell_{2} ℓ 1 < Φ ℓ 2 (or simply ℓ 1 < ℓ 2 \ell_{1}<\ell_{2} ℓ 1 < ℓ 2 if the formula is obvious) if lv ( ℓ 1 ) < lv ( ℓ 2 ) \operatorname{lv}(\ell_{1})<\operatorname{lv}(\ell_{2}) lv ( ℓ 1 ) < lv ( ℓ 2 ) .
Definition 42 (QDNF, AQBF). A QDNF is a QBF Q ⋅ ψ Q\cdot\psi Q ⋅ ψ such that ψ \psi ψ is a disjunction of cubes (DNF). An augmented QBF (AQBF) is a QBF of the form Φ = Q ⋅ ( φ ∨ ψ ) \Phi=Q\cdot(\varphi\lor\psi) Φ = Q ⋅ ( φ ∨ ψ ) , such that φ \varphi φ is a conjunction of clauses (CNF) and ψ \psi ψ is a disjunction of cubes (DNF). We denote C ( Φ ) : = φ \mathcal{C}(\Phi):=\varphi C ( Φ ) := φ and D ( Φ ) : = ψ \mathcal{D}(\Phi):=\psi D ( Φ ) := ψ .
Definition 43 ((Long-distance) Q-resolution and Q-consensus). Let C 1 C_{1} C 1 and C 2 C_{2} C 2 be two clauses (cubes). Let ℓ \ell ℓ be a literal.
The resolvent of C 1 ∨ ℓ C_{1}\lor\ell C 1 ∨ ℓ and C 2 ∨ ℓ ‾ C_{2}\lor\overline{\ell} C 2 ∨ ℓ over ℓ \ell ℓ is defined as
( C 1 ∨ ℓ ) ⋈ ℓ ( C 2 ∨ ℓ ‾ ) : = C 1 ∨ C 2 (C_{1}\lor\ell)\mathbin{\bowtie^{\ell}}(C_{2}\lor\overline{\ell}):=C_{1}\lor C_{2} ( C 1 ∨ ℓ ) ⋈ ℓ ( C 2 ∨ ℓ ) := C 1 ∨ C 2 (resp. ( C 1 ∧ ℓ ) ⋈ ℓ ( C 2 ∧ ℓ ‾ ) : = C 1 ∧ C 2 (C_{1}\land\ell)\mathbin{\bowtie^{\ell}}(C_{2}\land\overline{\ell}):=C_{1}\land C_{2} ( C 1 ∧ ℓ ) ⋈ ℓ ( C 2 ∧ ℓ ) := C 1 ∧ C 2 ).
Let C : = ℓ 1 ∨ … ∨ ℓ n C:=\ell_{1}\lor\ldots\lor\ell_{n} C := ℓ 1 ∨ … ∨ ℓ n be a clause from a QCNF or AQBF Φ \Phi Φ such that ℓ i ≤ Φ ℓ j \ell_{i}\leq_{\Phi}\ell_{j} ℓ i ≤ Φ ℓ j for all 1 ≤ i < j ≤ n 1\leq i<j\leq n 1 ≤ i < j ≤ n . Let p ∈ { 1 , … , n + 1 } p\in\{1,\ldots,n+1\} p ∈ { 1 , … , n + 1 } be minimal such that all of ℓ p , … , ℓ n \ell_{p},\ldots,\ell_{n} ℓ p , … , ℓ n are universal (so p = n + 1 p=n+1 p = n + 1 if ℓ n \ell_{n} ℓ n is existential). Then we can perform a universal reduction step and obtain
red Φ ∀ ( C ) : = ℓ 1 ∨ … ∨ ℓ p − 1 . \operatorname{red}^{\forall}_{\Phi}(C):=\ell_{1}\lor\ldots\lor\ell_{p-1}. red Φ ∀ ( C ) := ℓ 1 ∨ … ∨ ℓ p − 1 . Analogously, we perform existential reduction on cubes, which we denote as red Φ ∃ ( C ) \operatorname{red}^{\exists}_{\Phi}(C) red Φ ∃ ( C ) . If it is clear whether C C C is a clause or a cube, we can just write red Φ ( C ) \operatorname{red}_{\Phi}(C) red Φ ( C ) or even red ( C ) \operatorname{red}(C) red ( C ) , if the QBF Φ \Phi Φ is also obvious. We will write red ( Φ ) = red Φ ( Φ ) \operatorname{red}(\Phi)=\operatorname{red}_{\Phi}(\Phi) red ( Φ ) = red Φ ( Φ ) , if we reduce all clauses of the AQBF Φ \Phi Φ according to its prefix.
Long-distance resolution goes back to the solver of Zhang and Malik [41 ] ; Balabanov and Jiang [1 ] formalised it as the proof system LD-Q-Res (LD-Q-Con), which extends Q-Res (Q-Con) proofs by replacing the resolution rule by
Resolution (long-distance): C i = C j ⋈ Φ x C k C_{i}=C_{j}\bowtie^{x}_{\Phi}C_{k} C i = C j ⋈ Φ x C k with x x x existential (universal) and j , k < i j,k<i j , k < i . The resolvent C i C_{i} C i is allowed to contain tautologies such as u ∨ u ‾ u\lor\overline{u} u ∨ u (resp. contradictions u ∧ u ‾ u\land\overline{u} u ∧ u ). If there is such a universal (existential) variable u u u that occur in C j C_{j} C j and C k C_{k} C k with opposite polarities, then we require x ≠ u x\ne u x = u .
Definition 44 (trails, [8 ] ). A trail T T T for a QCNF or AQBF Φ \Phi Φ is a (finite) sequence of pairwise distinct literals from Φ \Phi Φ , including the empty literals ⊥ \bot ⊥ and ⊤ \top ⊤ . Each two literals in
T \mathcal{T} T have to correspond to pairwise distinct variables from Φ \Phi Φ . In general, a trail has the form
T = ( p ( 0 , 1 ) , … , p ( 0 , g 0 ) ; d 1 , p ( 1 , 1 ) , … , p ( 1 , g 1 ) ; … ; d r , p ( r , 1 ) , … , p ( r , g r ) ) , (4) \mathcal{T} = (p_{(0,1)},\ldots,p_{(0,g_0)};\mathbf{d}_1,p_{(1,1)},\ldots,p_{(1,g_1)};\ldots;\mathbf{d}_r,p_{(r,1)},\ldots,p_{(r,g_r)}), \tag*{(4)} T = ( p ( 0 , 1 ) , … , p ( 0 , g 0 ) ; d 1 , p ( 1 , 1 ) , … , p ( 1 , g 1 ) ; … ; d r , p ( r , 1 ) , … , p ( r , g r ) ) , (4)
where the d i d_i d i are decision literals and p ( i , j ) p_{(i,j)} p ( i , j ) are propagated literals. Decision literals are written in boldface . We use a semicolon before each decision to mark the end of a decision level. If one of the empty literals ⊥ \bot ⊥ or ⊤ \top ⊤ is contained in T \mathcal{T} T , then it has to be the last literal p ( r , g r ) p_{(r,g_r)} p ( r , g r ) . In this case, we say that T \mathcal{T} T has run into a conflict.
Trails can be interpreted as non-tautological sets of literals, and therefore as (partial) assignments. We define the length ∣ T ∣ |\mathcal{T}| ∣ T ∣ of a trail as the number of its literals, where we do not count the empty literals ⊥ \bot ⊥ and ⊤ \top ⊤ . We write x < T y x <_{\mathcal{T}} y x < T y if x , y ∈ T x,y \in\mathcal{T} x , y ∈ T and x x x is left of y y y in T \mathcal{T} T . Furthermore, we write x ≤ T y x \leq_{\mathcal{T}} y x ≤ T y if x < T y x <_{\mathcal{T}} y x < T y or x = y x=y x = y .
As trails are produced gradually from left to right in an algorithm, we define T [ i , j ] \mathcal{T}[i,j] T [ i , j ] for i ∈ { 0 , … , r } i \in\{0,\ldots,r\} i ∈ { 0 , … , r } and j ∈ { 0 , … , g i } j \in\{0,\ldots,g_i\} j ∈ { 0 , … , g i } as the subtrail that contains all literals from T \mathcal{T} T up to (and excluding) p ( i , j ) p_{(i,j)} p ( i , j ) (resp. d i d_i d i , if j = 0 j=0 j = 0 ) in the same order. Intuitively, T [ i , j ] \mathcal{T}[i,j] T [ i , j ] is the state of the trail before we assigned the literal at the point [ i , j ] [i,j] [ i , j ] (which is p ( i , j ) p_{(i,j)} p ( i , j ) or d i d_i d i ).
For each propagated literal p ( i , j ) p_{(i,j)} p ( i , j ) in a trail T \mathcal{T} T the formula must contain a clause or a cube that caused this propagation by becoming a unit clause/cube under the current trail assignment. We call such a clause (cube) the antecedent clause (cube) of p ( i , j ) p_{(i,j)} p ( i , j ) and denote it by ante T ( p ( i , j ) ) \operatorname{ante}_{\mathcal{T}}(p_{(i,j)}) ante T ( p ( i , j ) ) .
We call a trail T \mathcal{T} T natural for a formula Φ \Phi Φ , if for each i ∈ { 1 , … , r } i \in\{1,\ldots,r\} i ∈ { 1 , … , r } the formula red ( Φ ∣ T [ i , 0 ] ) \operatorname{red}(\Phi|_{\mathcal{T}[i,0]}) red ( Φ ∣ T [ i , 0 ] ) contains no unit or empty constraints. Furthermore, the formula red ( Φ ∣ T [ i , j ] ) \operatorname{red}(\Phi|_{\mathcal{T}[i,j]}) red ( Φ ∣ T [ i , j ] ) must not contain empty constraints for each i ∈ { 0 , … , r } i \in\{0,\ldots,r\} i ∈ { 0 , … , r } , j ∈ { 1 , … , g i } j \in\{1,\ldots,g_i\} j ∈ { 1 , … , g i } , except [ i , j ] = [ r , g r ] [i,j]=[r,g_r] [ i , j ] = [ r , g r ] . Intuitively, this means that decisions are only made if there are no more propagations possible on the same decision level, and conflicts have to be taken care of immediately. Note that the first condition says nothing about the empty trail T [ 0 , 0 ] \mathcal{T}[0,0] T [ 0 , 0 ] : the formula Φ \Phi Φ itself may contain unit clauses (for example learned unit clauses), and these are propagated on decision level 0.
Definition 45 (learnable constraints, [8 ] ). Let T \mathcal{T} T be a trail for Φ \Phi Φ of the form (2) with p ( r , g r ) ∈ { ⊥ , ⊤ } p_{(r,g_r)} \in\{\bot,\top\} p ( r , g r ) ∈ { ⊥ , ⊤ } . Starting with ante T ( ⊥ ) \operatorname{ante}_{\mathcal{T}}(\bot) ante T ( ⊥ ) (resp. ante T ( ⊤ ) \operatorname{ante}_{\mathcal{T}}(\top) ante T ( ⊤ ) ) we reversely resolve with the antecedent clauses (cubes) that were used to propagate the existential (universal) variables, until we stop at some point. Literals that were propagated via cubes (clauses) will be interpreted as decisions. If a resolution step cannot be performed at some point due to a missing pivot, we simply skip that antecedent. The clause (cube) we so derive, after universal (existential) reduction, is a learnable constraint. We denote the sequence of learnable constraints by L ( T ) \mathcal{L}(\mathcal{T}) L ( T ) .
We can also learn cubes from trails that did not run into conflict. If T \mathcal{T} T is a total assignment of the variables from Φ \Phi Φ , then we define the set of learnable constraints as the set of cubes
L ( T ) : = { red Φ ( D ) : D ⊆ T and D satisfies C ( Φ ) } . \mathcal{L}(\mathcal{T}) := \{\operatorname{red}_{\Phi}(D) : D \subseteq\mathcal{T}\text{ and }D\text{ satisfies }\mathcal{C}(\Phi)\}. L ( T ) := { red Φ ( D ) : D ⊆ T and D satisfies C ( Φ )} .
Definition 46 (QCDCL proof system, [8 ] ). A QCDCL proof ι \iota ι from a QCNF Φ = Q ⋅ ϕ \Phi= Q \cdot\phi Φ = Q ⋅ ϕ of a clause or cube C C C is a sequence of triples
ι : = [ ( T i , C i , π i ) ] i = 1 n , \iota:= [(\mathcal{T}_i,C_i,\pi_i)]_{i=1}^{n}, ι := [( T i , C i , π i ) ] i = 1 n , where C n = C C_n=C C n = C , each T i \mathcal{T}_i T i is a trail for the current formula Φ i \Phi_i Φ i (defined below), such that
for each decision d i d_i d i we have that ∣ Var Φ ∣ T [ i , 0 ] ( d i ) ∣ = 1 |\operatorname{Var}_{\Phi|_{\mathcal{T}[i,0]}}(d_i)|=1 ∣ Var Φ ∣ T [ i , 0 ] ( d i ) ∣ = 1 (i.e., decisions are level-ordered: only variables from the leftmost block that is not yet completely assigned may be decided),
reductions during unit propagation have to be performed all the time – formally, for each propagated literal p ( i , j ) ∈ T p_{(i,j)} \in\mathcal{T} p ( i , j ) ∈ T we need red ( ante T ( p ( i , j ) ) ∣ T [ i , j ] ) = ( p ( i , j ) ) \operatorname{red}(\operatorname{ante}_{\mathcal{T}}(p_{(i,j)})|_{\mathcal{T}[i,j]})=(p_{(i,j)}) red ( ante T ( p ( i , j ) ) ∣ T [ i , j ] ) = ( p ( i , j ) ) (or [ p ‾ ( i , j ) ] [\overline{p}_{(i,j)}] [ p ( i , j ) ] ),
unit clauses (cubes) can only propagate existential (universal) literals – formally, this means that if p ( i , j ) p_{(i,j)} p ( i , j ) is existential (universal), then ante T ( p ( i , j ) ) \operatorname{ante}_{\mathcal{T}}(p_{(i,j)}) ante T ( p ( i , j ) ) has to be a clause (cube),
each C i ∈ L ( T i ) C_i \in\mathcal{L}(T_i) C i ∈ L ( T i ) is one of the constraints we can learn from the trail, and π i \pi_i π i is the LD-Q-Res or LD-Q-Con derivation of C i C_i C i we get by performing the steps in Definition 45. We define R ( ι ) \mathcal{R}(\iota) R ( ι ) as the extracted proof of C C C that we get by sticking together suitable π i \pi_i π i .
We set Φ 1 : = Φ \Phi_1 := \Phi Φ 1 := Φ and obtain Φ j + 1 \Phi_{j+1} Φ j + 1 by adding the clause C j C_j C j to C ( Φ j ) C(\Phi_j) C ( Φ j ) (resp. adding the cube C j C_j C j to D ( Φ j ) D(\Phi_j) D ( Φ j ) ). We require that T 1 T_1 T 1 is a natural trail and for each 2 ≤ i ≤ n 2 \le i \le n 2 ≤ i ≤ n there is a point [ a i , b i ] [a_i,b_i] [ a i , b i ] such that T i [ a i , b i ] = T i − 1 [ a i , b i ] T_i[a_i,b_i] = T_{i-1}[a_i,b_i] T i [ a i , b i ] = T i − 1 [ a i , b i ] and T i \ T i [ a i , b i ] T_i \backslash T_i[a_i,b_i] T i \ T i [ a i , b i ] has to be a natural trail for Φ i ∣ T i [ a i , b i ] \Phi_i|_{T_i[a_i,b_i]} Φ i ∣ T i [ a i , b i ] . This process is called backtracking. If T i − 1 [ a i , b i ] = ∅ T_{i-1}[a_i,b_i] = \emptyset T i − 1 [ a i , b i ] = ∅ , then this is also called a restart.
For C = ( ⊥ ) C = (\bot) C = ( ⊥ ) , ι \iota ι is a QCDCL refutation of Φ \Phi Φ . For C = [ ⊤ ] C = [\top] C = [ ⊤ ] , ι \iota ι is a QCDCL verification of Φ \Phi Φ . The trail size of ι \iota ι is defined as ∣ ι ∣ : = ∑ i = 1 n ∣ T i ∣ |\iota| := \sum_{i=1}^{n} |T_i| ∣ ι ∣ := ∑ i = 1 n ∣ T i ∣ . Note that all QCDCL refutations constructed in this paper only learn clauses. In particular, they do not learn cubes from total assignments.
Definition 47 (CDCL steps). Let ι = [ ( T i , C i , π i ) ] i = 1 n \iota= [(T_i, C_i, \pi_i)]_{i=1}^{n} ι = [( T i , C i , π i ) ] i = 1 n be a QCDCL refutation of a false QCNF Φ \Phi Φ . If there exist an existential quantifier block ∃ Y \exists Y ∃ Y of Φ \Phi Φ and a point [ a , b ] [a,b] [ a , b ] in some T j T_j T j such that
all variables left of Y Y Y are already assigned in T j T_j T j prior to [ a , b ] [a,b] [ a , b ] ,
⋀ C ∈ red ( Φ j ∣ T j [ a , b ] ) , var ( C ) ⊆ Y C \bigwedge_{C \in\operatorname{red}(\Phi_j|_{T_j[a,b]}),\operatorname{var}(C) \subseteq Y} C ⋀ C ∈ red ( Φ j ∣ T j [ a , b ] ) , var ( C ) ⊆ Y C is an unsatisfiable propositional formula (where red denotes the universal reduction of all clauses, which is performed anyway during unit propagation),
after [ a , b ] [a,b] [ a , b ] only variables from Y Y Y get assigned,
the point [ a , b ] [a,b] [ a , b ] was not overstepped, neither by the learning of C j C_j C j , nor by backtracking from T j T_j T j to T j + 1 T_{j+1} T j + 1 — that means the derivation π j \pi_j π j of C j C_j C j only uses antecedent clauses of literals that were assigned after [ a , b ] [a,b] [ a , b ] , and T j + 1 T_{j+1} T j + 1 starts with T j [ a , b ] T_j[a,b] T j [ a , b ] (if j < n j < n j < n ),
then ( T j , C j , π j ) (T_j, C_j, \pi_j) ( T j , C j , π j ) is called a CDCL step in ι \iota ι . For P : = { j : ( T j , C j , π j ) is a CDCL step } P := \{j : (T_j, C_j, \pi_j)\text{ is a CDCL step}\} P := { j : ( T j , C j , π j ) is a CDCL step } , we define ∣ ι ∣ C D C L : = ∑ j ∈ P ∣ T j ∣ |\iota|_{\mathrm{CDCL}} := \sum_{j \in P} |T_j| ∣ ι ∣ CDCL := ∑ j ∈ P ∣ T j ∣ . Hence ∣ ι ∣ − ∣ ι ∣ C D C L |\iota| - |\iota|_{\mathrm{CDCL}} ∣ ι ∣ − ∣ ι ∣ CDCL measures “genuine” QCDCL size, discarding pure CDCL runs.
To ensure that such CDCL steps are cleanly detectable, we require that trails follow a certain structure. A way to achieve this is by imposing the XT-property on the QCNF [15 ] . Informally, this XT-property causes the trails to be fully level-ordered (from X X X - to U U U - to T T T -variables), which is important so that propagations do not get mixed up too much. Originally, the XT-property was designed to obtain QCDCL lower bounds (cf. [15 ] ) as level-ordered trails turn out to be disadvantageous for QCDCL runtime. Ironically, this XT-property will now help us to obtain upper bounds for QCDCL.
Definition 48 (XT-property, [15 ] ). A Σ 3 b \Sigma^b_3 Σ 3 b QCNF with the prefix ∃ X ∀ U ∃ T \exists X \forall U \exists T ∃ X ∀ U ∃ T fulfils the XT-property if no two clauses C 1 , C 2 C_1,C_2 C 1 , C 2 with var ( C 1 ) , var ( C 2 ) ⊆ T \operatorname{var}(C_1),\operatorname{var}(C_2) \subseteq T var ( C 1 ) , var ( C 2 ) ⊆ T (“T T T -clauses”) are resolvable and if it contains no unit clauses ℓ \ell ℓ with ℓ ∈ T \ell\in T ℓ ∈ T (“unit T T T -clauses”) and no clauses C C C with var ( C ) ∩ X ≠ ∅ \operatorname{var}(C) \cap X \ne\emptyset var ( C ) ∩ X = ∅ , var ( C ) ∩ T ≠ ∅ \operatorname{var}(C) \cap T \ne\emptyset var ( C ) ∩ T = ∅ and var ( C ) ∩ U = ∅ \operatorname{var}(C) \cap U = \emptyset var ( C ) ∩ U = ∅ (“XT-clauses”).
Note that the XT-property still holds after learning a clause (cf. [15 ] ). This means that a formula with the XT-property will continue to fulfil it for the entire QCDCL run. We can now show:
Theorem 49. Let Φ \Phi Φ be a false Σ 3 b \Sigma^b_3 Σ 3 b QCNF that fulfils the XT-property. Then there exists a QCDCL refutation ι \iota ι with ∣ ι ∣ − ∣ ι ∣ C D C L ≤ cw − ( Φ ) ⋅ ∣ var ( Φ ) ∣ |\iota| - |\iota|_{\mathrm{CDCL}} \le\operatorname{cw}^{-}(\Phi) \cdot|\operatorname{var}(\Phi)| ∣ ι ∣ − ∣ ι ∣ CDCL ≤ cw − ( Φ ) ⋅ ∣ var ( Φ ) ∣ .
Proof. Let Φ = ∃ X ∀ U ∃ T ⋅ C 1 ∧ ⋯ ∧ C k \Phi= \exists X \forall U \exists T \cdot C_1 \land\cdots\land C_k Φ = ∃ X ∀ U ∃ T ⋅ C 1 ∧ ⋯ ∧ C k . We can assume that no clause of Φ \Phi Φ consists of universal literals only. Otherwise red ( Φ ) \operatorname{red}(\Phi) red ( Φ ) contains the empty clause, and the trail ( ⊥ ) (\bot) ( ⊥ ) alone is a QCDCL refutation ι \iota ι with ∣ ι ∣ = 0 |\iota| = 0 ∣ ι ∣ = 0 . We call a clause T T T -free if it contains no literal over T T T . For a T T T -free clause C C C we have red ( C ) = C ∣ X \operatorname{red}(C) = C|_X red ( C ) = C ∣ X , because all universal literals of C C C are quantified right of its existential literals.
Let us first collect what we need from the XT-property. Since the XT-property is preserved by clause learning, it holds for all formulas Φ j \Phi_j Φ j that occur during the run we construct. In particular, every clause of Φ j \Phi_j Φ j that contains a literal over T T T either contains a literal over U U U as well, or it is a T T T -clause with at least two literals. Hence, as long as no variable from T T T is assigned, such a clause can neither become unit nor empty. That means while only variables from X X X are assigned, unit propagation only assigns literals over X X X , and the antecedent clauses and conflict clauses are T T T -free. (P)
Now let
M = ( D 1 , μ 1 ) , … , ( D z , μ z ) \mathcal{M}=(D_1,\mu_1),\ldots,(D_z,\mu_z) M = ( D 1 , μ 1 ) , … , ( D z , μ z )
be the UDL that we have constructed in the proof of Theorem 37 from a universal winning strategy s s s . Let us recall this construction for the case of a Σ 3 3 \Sigma^3_3 Σ 3 3 formula, where m = 1 m=1 m = 1 and U 0 = ∅ U_0=\varnothing U 0 = ∅ . Here we have game ( s , σ , X ) = σ \operatorname{game}(s,\sigma,X)=\sigma game ( s , σ , X ) = σ for every σ ∈ ⟨ X ⟩ \sigma\in\langle X\rangle σ ∈ ⟨ X ⟩ , so the optimal chains with respect to s s s are exactly the optimal moves σ ∈ ⟨ X ⟩ \sigma\in\langle X\rangle σ ∈ ⟨ X ⟩ of the block ∃ X \exists X ∃ X , and the fruition of σ \sigma σ is ( sat ( σ ) ) (\operatorname{sat}(\sigma)) ( sat ( σ )) . The elements γ 1 < ⋯ < γ z \gamma_1<\cdots<\gamma_z γ 1 < ⋯ < γ z of the set H H H from that proof are therefore optimal moves in ⟨ X ⟩ \langle X\rangle ⟨ X ⟩ . They have pairwise different sets I i : = sat ( γ i ) I_i:=\operatorname{sat}(\gamma_i) I i := sat ( γ i ) , as shown there, and we have ξ i = ( I i ) \xi_i=(I_i) ξ i = ( I i ) . Furthermore,
D i = ⋀ r ∉ I i ¬ ( C r ∣ X ) and μ i = s ( γ i ) D_i=\bigwedge_{r\notin I_i}\neg(C_r|_X)
\qquad\text{and}\qquad
\mu_i=s(\gamma_i) D i = r ∈ / I i ⋀ ¬ ( C r ∣ X ) and μ i = s ( γ i )
for i = 1 , … , z i=1,\ldots,z i = 1 , … , z . By Theorem 37 we have z ≤ cw − ( Φ ) m = cw − ( Φ ) z\leq\operatorname{cw}^{-}(\Phi)^m=\operatorname{cw}^{-}(\Phi) z ≤ cw − ( Φ ) m = cw − ( Φ ) . Because M \mathcal{M} M is a UDL, we know that ⋁ i = 1 z D i ≡ ⊤ \bigvee_{i=1}^{z}D_i\equiv\top ⋁ i = 1 z D i ≡ ⊤ , hence ⋀ i = 1 z ¬ D i \bigwedge_{i=1}^{z}\neg D_i ⋀ i = 1 z ¬ D i is unsatisfiable. Note that ¬ D i \neg D_i ¬ D i is a clause over X X X , and that C r ∣ X ⊆ ¬ D i C_r|_X\subseteq\neg D_i C r ∣ X ⊆ ¬ D i for every r ∉ I i r\notin I_i r ∈ / I i .
The plan is as follows. We construct the refutation ι \iota ι in z z z stages. In stage i i i we make sure that the current formula contains a T T T -free clause E i E_i E i with red ( E i ) ⊆ ¬ D i \operatorname{red}(E_i)\subseteq\neg D_i red ( E i ) ⊆ ¬ D i , and we use at most one trail that is not a CDCL step for this. Once all these clauses are present, the formula ⋀ i = 1 z red ( E i ) \bigwedge_{i=1}^{z}\operatorname{red}(E_i) ⋀ i = 1 z red ( E i ) is an unsatisfiable CNF over X X X , and we can learn the empty clause with CDCL steps. During the construction we maintain the following invariant for every learned clause L L L :
(*) There is a nonempty set R ( L ) ⊆ { 1 , … , k } R(L)\subseteq\{1,\ldots,k\} R ( L ) ⊆ { 1 , … , k } such that L ∣ X = ⋃ r ∈ R ( L ) C r ∣ X L|_X=\bigcup_{r\in R(L)}C_r|_X L ∣ X = ⋃ r ∈ R ( L ) C r ∣ X . In other words, L L L contains all X X X -literals of the original clauses it stems from, and no other X X X -literals. For an original clause we set R ( C r ) : = { r } R(C_r):=\{r\} R ( C r ) := { r } . The invariant is preserved as long as the pivots of the resolution steps are variables from T T T : X X X -literals are never removed by such resolution steps, and they are never removed by universal reduction.
The crucial consequence of () is the following. Let L L L be an original or learned clause that fulfils ( ), and assume that γ i \gamma_i γ i falsifies all literals of L ∣ X L|_X L ∣ X . Then γ i \gamma_i γ i falsifies C r ∣ X C_r|_X C r ∣ X for every r ∈ R ( L ) r\in R(L) r ∈ R ( L ) . As γ i \gamma_i γ i only assigns variables from X X X , it does not satisfy C r C_r C r , and we get r ∉ sat ( γ i ) = I i r\notin\operatorname{sat}(\gamma_i)=I_i r ∈ / sat ( γ i ) = I i . Therefore, D i D_i D i contains ¬ ( C r ∣ X ) \neg(C_r|_X) ¬ ( C r ∣ X ) , which means C r ∣ X ⊆ ¬ D i C_r|_X\subseteq\neg D_i C r ∣ X ⊆ ¬ D i . We conclude
L ∣ X ⊆ ¬ D i whenever L fulfils (*) and γ i falsifies all literals of L ∣ X . (5) L|_X\subseteq\neg D_i\qquad\text{whenever }L\text{ fulfils (*) and }\gamma_i\text{ falsifies all literals of }L|_X. \tag*{(5)} L ∣ X ⊆ ¬ D i whenever L fulfils (*) and γ i falsifies all literals of L ∣ X . (5)
We now describe stage i i i . Let Φ j \Phi_j Φ j be the current formula, that is, Φ \Phi Φ together with the clauses learned so far, all of which fulfil (*). If the empty clause has already been learned, we are done. Otherwise we build a natural trail as follows. We perform all possible unit propagations. Whenever no propagation is possible and X X X is not yet completely assigned, we decide the next unassigned variable of X X X with the value that γ i \gamma_i γ i gives to it. By (P), only literals over X X X are propagated in this phase, and only T T T -free clauses are involved. We say that the trail is consistent with γ i \gamma_i γ i if all its literals belong to γ i \gamma_i γ i . We distinguish two cases. In the first case, a conflict occurs or a literal ℓ \ell ℓ with ℓ ˉ ∈ γ i \bar{\ell} \in\gamma_i ℓ ˉ ∈ γ i is propagated at some point while the trail is still consistent with γ i \gamma_i γ i . Let C C C be the conflict clause (resp. the antecedent clause of ℓ \ell ℓ ). By (P), C C C is T T T -free and red ( C ) = C ∣ X \operatorname{red}(C) = C|_X red ( C ) = C ∣ X . All literals of C ∣ X C|_X C ∣ X are falsified by γ i \gamma_i γ i : the literals other than ℓ \ell ℓ are falsified by the trail, which is consistent with γ i \gamma_i γ i , and we have ℓ ˉ ∈ γ i \bar{\ell} \in\gamma_i ℓ ˉ ∈ γ i . Since C C C fulfils ( ∗ ) (*) ( ∗ ) , (3) gives us red ( C ) = C ∣ X ⊆ ¬ D i \operatorname{red}(C) = C|_X \subseteq\neg D_i red ( C ) = C ∣ X ⊆ ¬ D i . Hence, the current formula already contains a clause with the required property, and we set E i : = C E_i := C E i := C . In this case we do not add any trail to ι \iota ι – the trail we have considered only served to find the clause E i E_i E i .
In the second case, no such event occurs. Then the trail assigns all variables of X X X according to γ i \gamma_i γ i , no other variables are assigned, and no conflict has occurred. We continue the trail by deciding the variables of U U U according to μ i \mu_i μ i , one after the other, each time followed by all possible propagations, until a conflict occurs or all of U U U is assigned. Universal literals are never propagated from clauses, so the universal part of the trail is contained in μ i \mu_i μ i . Assume first that all of U U U gets assigned without a conflict, and let P \mathcal{P} P be the trail at this point. The strategy s s s is winning, hence Φ ∣ γ i ∪ μ i \Phi|_{\gamma_i\cup\mu_i} Φ ∣ γ i ∪ μ i is an unsatisfiable CNF over T T T . The clauses of red ( Φ j ∣ P ) \operatorname{red}(\Phi_j|_{\mathcal{P}}) red ( Φ j ∣ P ) over T T T therefore form an unsatisfiable propositional CNF Ψ \Psi Ψ as well: they include all clauses of Φ ∣ γ i ∪ μ i \Phi|_{\gamma_i\cup\mu_i} Φ ∣ γ i ∪ μ i restricted by the literals over T T T that were propagated in P \mathcal{P} P . That means the point [ a , b ] [a,b] [ a , b ] at the end of P \mathcal{P} P fulfils the first two conditions of Definition 47 with Y : = T Y := T Y := T . We now run CDCL on Ψ \Psi Ψ : we decide variables from T T T and propagate until a conflict occurs; we learn a clause by resolving the conflict clause with the antecedent clauses of the literals over T T T that were propagated after [ a , b ] [a,b] [ a , b ] , in reverse order of the trail, until the clause is asserting with respect to the decision levels after [ a , b ] [a,b] [ a , b ] ; and we backtrack to the end of P \mathcal{P} P or to a later decision level. Hence [ a , b ] [a,b] [ a , b ] is never overstepped and all these trails are CDCL steps. Note that the learned clauses contain literals over X X X and U U U as well, namely the X X X - and U U U -literals of the clauses involved, all of which are falsified by γ i ∪ μ i \gamma_i\cup\mu_i γ i ∪ μ i . In particular, no complementary universal literals occur in these resolution steps, and the invariant ( ∗ ) (*) ( ∗ ) is preserved because all pivots are variables from T T T . Since Ψ \Psi Ψ is unsatisfiable, we arrive after finitely many CDCL steps at a trail T ∗ \mathcal{T}^* T ∗ in which a conflict occurs before any variable from T T T is decided, that is, on a universal decision level. If a conflict already occurs while we decide the variables of U U U , we let T ∗ \mathcal{T}^* T ∗ be this trail, and there are no CDCL steps in this stage at all.
In T ∗ \mathcal{T}^* T ∗ , all literals over T T T are propagated. Let C C C be the conflict clause of T ∗ \mathcal{T}^* T ∗ . We now resolve C C C with the antecedent clauses of all propagated literals over T T T in reverse order of the trail. Each antecedent clause has all its X X X -literals falsified by γ i \gamma_i γ i and all its U U U -literals falsified by μ i \mu_i μ i – its universal literals cannot be reduced, since the clause contains a literal over T T T . The conflict clause C C C also has all its X X X -literals falsified by γ i \gamma_i γ i , and each of its U U U -literals is falsified by μ i \mu_i μ i or unassigned (in the latter case it was removed by reduction). Hence no complementary universal literals occur in these resolution steps, and we obtain a clause over X ∪ U X \cup U X ∪ U . Universal reduction removes all its literals over U U U , and we learn the resulting clause E i E_i E i over X X X , whose literals are all falsified by γ i \gamma_i γ i . This is a learnable constraint in the sense of Definition 45: we stop the reverse resolution before we reach the antecedents of the literals over X X X . The clause E i E_i E i fulfils ( ∗ ) (*) ( ∗ ) , because all pivots were variables from T T T . By (3), we get E i ⊆ ¬ D i E_i \subseteq\neg D_i E i ⊆ ¬ D i . After learning E i E_i E i , we restart, that is, the next trail begins with the empty trail. The trail T ∗ \mathcal{T}^* T ∗ is the only trail of this stage that is possibly not a CDCL step.
After stage z z z , the current formula contains T T T -free clauses E 1 , … , E z E_1,\ldots,E_z E 1 , … , E z with red ( E i ) ⊆ ¬ D i \operatorname{red}(E_i) \subseteq\neg D_i red ( E i ) ⊆ ¬ D i for i = 1 , … , z i=1,\ldots,z i = 1 , … , z . As ⋀ i = 1 z ¬ D i \bigwedge_{i=1}^{z}\neg D_i ⋀ i = 1 z ¬ D i is unsatisfiable, the CNF ⋀ i = 1 z red ( E i ) \bigwedge_{i=1}^{z}\operatorname{red}(E_i) ⋀ i = 1 z red ( E i ) over X X X is unsatisfiable, and so is the CNF that consists of all clauses of red ( Φ j ) \operatorname{red}(\Phi_j) red ( Φ j ) over X X X . We refute it by CDCL steps V 1 , … , V c \mathcal{V}_1,\ldots,\mathcal{V}_c V 1 , … , V c for some c ∈ N c \in\mathbb{N} c ∈ N : we decide and propagate variables from X X X only – by (P) no other variables get assigned – and learn clauses by resolving over variables from X X X . Complementary universal literals may occur in these resolution steps, for instance when we resolve x ∨ u x \lor u x ∨ u with x ˉ ∨ u ˉ \bar{x} \lor\bar{u} x ˉ ∨ u ˉ , but this is allowed by the long-distance resolution rule, as all universal variables are quantified right of X X X , and the universal literals are removed by reduction afterwards. These trails are CDCL steps with Y : = X Y := X Y := X and the point [ a , b ] = [ 0 , 0 ] [a,b] = [0,0] [ a , b ] = [ 0 , 0 ] at the beginning of the trail: the first condition of Definition 47 is void, the second one holds since the clauses of red ( Φ j ) \operatorname{red}(\Phi_j) red ( Φ j ) over X X X are unsatisfiable, and the beginning of the trail is never overstepped.
All in all, ι \iota ι consists of the trails of the stages 1 , … , z 1,\ldots,z 1 , … , z followed by V 1 , … , V c \mathcal{V}_1,\ldots,\mathcal{V}_c V 1 , … , V c , and the only trails that are not CDCL steps are the trails T ∗ T^* T ∗ of the stages that fall into the second case. There are at most z ≤ cw − ( Φ ) z \le\operatorname{cw}^{-}(\Phi) z ≤ cw − ( Φ ) of them. Each trail assigns every variable at most once, hence it has length at most ∣ var ( Φ ) ∣ |\operatorname{var}(\Phi)| ∣ var ( Φ ) ∣ (recall that ⊥ \bot ⊥ is not counted). We conclude that ∣ ι ∣ − ∣ ι ∣ C D C L ≤ cw − ( Φ ) ⋅ ∣ var ( Φ ) ∣ |\iota| - |\iota|_{\mathrm{CDCL}} \le\operatorname{cw}^{-}(\Phi) \cdot|\operatorname{var}(\Phi)| ∣ ι ∣ − ∣ ι ∣ CDCL ≤ cw − ( Φ ) ⋅ ∣ var ( Φ ) ∣ .
Corollary 51. QCDCL simulates CAQEC o r e ^{Core} C or e on false Σ 3 b \Sigma^b_3 Σ 3 b QCNFs that fulfil the XT-property, under the condition that CDCL steps in QCDCL are not counted. More precisely, every CAQEC o r e ^{Core} C or e refutation κ \kappa κ of such a QCNF Φ \Phi Φ can be turned into a QCDCL refutation ι \iota ι of Φ \Phi Φ with ∣ ι ∣ − ∣ ι ∣ C D C L ≤ ∣ κ ∣ ⋅ ∣ var ( Φ ) ∣ |\iota| - |\iota|_{\mathrm{CDCL}} \le|\kappa| \cdot|\operatorname{var}(\Phi)| ∣ ι ∣ − ∣ ι ∣ CDCL ≤ ∣ κ ∣ ⋅ ∣ var ( Φ ) ∣ . By adding the preprocessing step as described in Remark 50, QCDCL even simulates CAQEC o r e ^{Core} C or e on all false Σ 3 b \Sigma^b_3 Σ 3 b QCNFs. In this case, ι \iota ι is a QCDCL refutation of the preprocessed QCNF Φ ′ \Phi' Φ ′ with ∣ ι ∣ − ∣ ι ∣ C D C L ≤ ∣ κ ∣ ⋅ ( ∣ var ( Φ ) ∣ + 1 ) |\iota| - |\iota|_{\mathrm{CDCL}} \le|\kappa| \cdot(|\operatorname{var}(\Phi)| + 1) ∣ ι ∣ − ∣ ι ∣ CDCL ≤ ∣ κ ∣ ⋅ ( ∣ var ( Φ ) ∣ + 1 ) .
Proof. By Proposition 32 , we have cw − ( Φ ) ≤ ∣ κ ∣ \operatorname{cw}^{-}(\Phi) \le|\kappa| cw − ( Φ ) ≤ ∣ κ ∣ , and the first bound follows from Theorem 49 . For the second statement, Remark 50 gives us cw − ( Φ ′ ) = cw − ( Φ ) ≤ ∣ κ ∣ \operatorname{cw}^{-}(\Phi') = \operatorname{cw}^{-}(\Phi) \le|\kappa| cw − ( Φ ′ ) = cw − ( Φ ) ≤ ∣ κ ∣ and ∣ var ( Φ ′ ) ∣ = ∣ var ( Φ ) ∣ + 1 |\operatorname{var}(\Phi')| = |\operatorname{var}(\Phi)| + 1 ∣ var ( Φ ′ ) ∣ = ∣ var ( Φ ) ∣ + 1 , so we can apply Theorem 49 to Φ ′ \Phi' Φ ′ .
As in Corollary 39 , this is a simulation in terms of proof size. The refutation ι \iota ι is constructed from the UDL of Theorem 37 , which we obtain from an optimal strategy, and we do not claim that ι \iota ι can be computed from κ \kappa κ in polynomial time.
We conclude this section with an example that illustrates the construction in the proof of Theorem 49 .
Example 52. Let Φ \Phi Φ be the QCNF from Example 2 with prefix ∃ x 1 , x 2 ∀ u 1 , u 2 ∃ y 1 , y 2 \exists x_1,x_2\forall u_1,u_2\exists y_1,y_2 ∃ x 1 , x 2 ∀ u 1 , u 2 ∃ y 1 , y 2 and the matrix
C 1 = x 1 ∨ u 1 ∨ y 1 C 2 = x 1 ∨ x 2 ∨ u 2 ∨ y 2 C 3 = y ˉ 1 ∨ y ˉ 2 C 4 = x ˉ 1 ∨ u ˉ 1 ∨ y 1 C 5 = x ˉ 1 ∨ u ˉ 2 ∨ y 2 C 6 = x ˉ 2 ∨ u ˉ 2 ∨ y 2 . \begin{aligned}
C_1 &= x_1 \lor u_1 \lor y_1 & C_2 &= x_1 \lor x_2 \lor u_2 \lor y_2 & C_3 &= \bar{y}_1 \lor\bar{y}_2 \\
C_4 &= \bar{x}_1 \lor\bar{u}_1 \lor y_1 & C_5 &= \bar{x}_1 \lor\bar{u}_2 \lor y_2 & C_6 &= \bar{x}_2 \lor\bar{u}_2 \lor y_2.
\end{aligned} C 1 C 4 = x 1 ∨ u 1 ∨ y 1 = x ˉ 1 ∨ u ˉ 1 ∨ y 1 C 2 C 5 = x 1 ∨ x 2 ∨ u 2 ∨ y 2 = x ˉ 1 ∨ u ˉ 2 ∨ y 2 C 3 C 6 = y ˉ 1 ∨ y ˉ 2 = x ˉ 2 ∨ u ˉ 2 ∨ y 2 . Note that Φ \Phi Φ fulfils the XT-property: there are no XT-clauses, and the only T-clause is C 3 C_3 C 3 , which is not a unit clause. Let s s s be the winning strategy for Φ \Phi Φ that sets s ( γ ) : = ( u 1 ↦ ( x 1 ) ∣ γ , u 2 ↦ ( x 1 ∨ x 2 ) ∣ γ ) s(\gamma) := (u_1 \mapsto(x_1)|_\gamma,u_2 \mapsto(x_1 \lor x_2)|_\gamma) s ( γ ) := ( u 1 ↦ ( x 1 ) ∣ γ , u 2 ↦ ( x 1 ∨ x 2 ) ∣ γ ) for γ ∈ ⟨ x 1 , x 2 ⟩ \gamma\in\langle x_1,x_2\rangle γ ∈ ⟨ x 1 , x 2 ⟩ . We now want to construct a UDL M \mathcal{M} M of size at most cw − ( Φ ) = 3 \operatorname{cw}^{-}(\Phi) = 3 cw − ( Φ ) = 3 and a QCDCL refutation ι \iota ι for Φ \Phi Φ with ∣ ι ∣ − ∣ ι ∣ C D C L ≤ cw − ( Φ ) ⋅ ∣ var ( Φ ) ∣ = 3 ⋅ 6 = 18 |\iota| - |\iota|_{\mathrm{CDCL}} \le\operatorname{cw}^{-}(\Phi) \cdot|\operatorname{var}(\Phi)| = 3 \cdot6 = 18 ∣ ι ∣ − ∣ ι ∣ CDCL ≤ cw − ( Φ ) ⋅ ∣ var ( Φ ) ∣ = 3 ⋅ 6 = 18 .
We start with computing the set of all optimal chains K ( Φ , s ) K(\Phi,s) K ( Φ , s ) . Since Φ \Phi Φ is a Σ 3 b \Sigma^b_3 Σ 3 b formula, K ( Φ , s ) K(\Phi,s) K ( Φ , s ) does not depend on s s s and only consists of optimal moves on the first quantifier block ∃ x 1 , x 2 \exists x_1,x_2 ∃ x 1 , x 2 . In Example 2 we have already computed all optimal moves for Φ \Phi Φ , hence we
obtain K ( Φ , s ) = { γ 1 , γ 2 , γ 3 } K(\Phi, s) = \{\gamma_1, \gamma_2, \gamma_3\} K ( Φ , s ) = { γ 1 , γ 2 , γ 3 } with γ 1 = x ˉ 1 ∧ x ˉ 2 \gamma_1 = \bar{x}_1 \land\bar{x}_2 γ 1 = x ˉ 1 ∧ x ˉ 2 , γ 2 = x ˉ 1 ∧ x 2 \gamma_2 = \bar{x}_1 \land x_2 γ 2 = x ˉ 1 ∧ x 2 and γ 3 = x 1 ∧ x ˉ 2 \gamma_3 = x_1 \land\bar{x}_2 γ 3 = x 1 ∧ x ˉ 2 . Next we compute their fruitions, which are fru ( γ 1 ) = ( { 4 , 5 , 6 } ) = : ξ 1 \operatorname{fru}(\gamma_1) = (\{4,5,6\}) =: \xi_1 fru ( γ 1 ) = ({ 4 , 5 , 6 }) =: ξ 1 , because γ 1 \gamma_1 γ 1 satisfies precisely C 4 C_4 C 4 , C 5 C_5 C 5 and C 6 C_6 C 6 , and fru ( γ 2 ) = ( { 2 , 4 , 5 } ) = : ξ 2 \operatorname{fru}(\gamma_2) = (\{2,4,5\}) =: \xi_2 fru ( γ 2 ) = ({ 2 , 4 , 5 }) =: ξ 2 and fru ( γ 3 ) = ( { 1 , 2 , 6 } ) = : ξ 3 \operatorname{fru}(\gamma_3) = (\{1,2,6\}) =: \xi_3 fru ( γ 3 ) = ({ 1 , 2 , 6 }) =: ξ 3 . Because Φ \Phi Φ is a Σ 3 \Sigma_3 Σ 3 formula, each fruition only consists of one set of indices. Note that the fruition of the assignment x 1 ∧ x 2 x_1 \land x_2 x 1 ∧ x 2 is ( { 1 , 2 } ) (\{1,2\}) ({ 1 , 2 }) , which is non-optimal because { 1 , 2 } \{1,2\} { 1 , 2 } is a proper subset of { 1 , 2 , 6 } \{1,2,6\} { 1 , 2 , 6 } . We conclude F ( Φ , s ) = { ξ 1 , ξ 2 , ξ 3 } F(\Phi, s) = \{\xi_1,\xi_2,\xi_3\} F ( Φ , s ) = { ξ 1 , ξ 2 , ξ 3 } . In this example, we have as many optimal fruitions as optimal moves. In general, we might have ∣ K ( Φ , s ) ∣ > ∣ F ( Φ , s ) ∣ |K(\Phi,s)| > |F(\Phi,s)| ∣ K ( Φ , s ) ∣ > ∣ F ( Φ , s ) ∣ because different optimal chains may have the same fruitions.
Now we can construct the cubes for our UDL by negating the clauses whose indices do not appear in the corresponding fruition, hence
D 1 = ⋀ r ∉ { 4 , 5 , 6 } ¬ C r ∣ X = ( ¬ C 1 ∧ ¬ C 2 ∧ ¬ C 3 ) ∣ X = x ˉ 1 ∧ x ˉ 2 D_1 = \bigwedge_{r\notin\{4,5,6\}} \neg C_r|_X = (\neg C_1 \land\neg C_2 \land\neg C_3)|_X = \bar{x}_1 \land\bar{x}_2 D 1 = r ∈ / { 4 , 5 , 6 } ⋀ ¬ C r ∣ X = ( ¬ C 1 ∧ ¬ C 2 ∧ ¬ C 3 ) ∣ X = x ˉ 1 ∧ x ˉ 2
D 2 = ⋀ r ∉ { 2 , 4 , 5 } ¬ C r ∣ X = ( ¬ C 1 ∧ ¬ C 3 ∧ ¬ C 6 ) ∣ X = x ˉ 1 ∧ x 2 D_2 = \bigwedge_{r\notin\{2,4,5\}} \neg C_r|_X = (\neg C_1 \land\neg C_3 \land\neg C_6)|_X = \bar{x}_1 \land x_2 D 2 = r ∈ / { 2 , 4 , 5 } ⋀ ¬ C r ∣ X = ( ¬ C 1 ∧ ¬ C 3 ∧ ¬ C 6 ) ∣ X = x ˉ 1 ∧ x 2
D 3 = ⋀ r ∉ { 1 , 2 , 6 } ¬ C r ∣ X = ( ¬ C 3 ∧ ¬ C 4 ∧ ¬ C 5 ) ∣ X = [ x 1 ] . D_3 = \bigwedge_{r\notin\{1,2,6\}} \neg C_r|_X = (\neg C_3 \land\neg C_4 \land\neg C_5)|_X = [x_1]. D 3 = r ∈ / { 1 , 2 , 6 } ⋀ ¬ C r ∣ X = ( ¬ C 3 ∧ ¬ C 4 ∧ ¬ C 5 ) ∣ X = [ x 1 ] .
We define the assignments μ 1 , μ 2 , μ 3 ∈ ⟨ u 1 , u 2 ⟩ \mu_1, \mu_2, \mu_3 \in\langle u_1,u_2\rangle μ 1 , μ 2 , μ 3 ∈ ⟨ u 1 , u 2 ⟩ according to the strategy s s s , hence μ 1 = u ˉ 1 ∧ u ˉ 2 \mu_1 = \bar{u}_1 \land\bar{u}_2 μ 1 = u ˉ 1 ∧ u ˉ 2 , μ 2 = u ˉ 1 ∧ u 2 \mu_2 = \bar{u}_1 \land u_2 μ 2 = u ˉ 1 ∧ u 2 and μ 3 = u 1 ∧ u 2 \mu_3 = u_1 \land u_2 μ 3 = u 1 ∧ u 2 . We do not need to alter s s s since it is the only winning strategy for Φ \Phi Φ anyway. It is easy to see that
M = ( D 1 , μ 1 ) , ( D 2 , μ 2 ) , ( D 3 , μ 3 ) = ( x ˉ 1 ∧ x ˉ 2 , u ˉ 1 ∧ u ˉ 2 ) , ( x ˉ 1 ∧ x 2 , u ˉ 1 ∧ u 2 ) , ( x 1 , u 1 ∧ u 2 ) \mathcal{M} = (D_1,\mu_1),(D_2,\mu_2),(D_3,\mu_3) = (\bar{x}_1 \land\bar{x}_2,\bar{u}_1 \land\bar{u}_2),(\bar{x}_1 \land x_2,\bar{u}_1 \land u_2),(x_1,u_1 \land u_2) M = ( D 1 , μ 1 ) , ( D 2 , μ 2 ) , ( D 3 , μ 3 ) = ( x ˉ 1 ∧ x ˉ 2 , u ˉ 1 ∧ u ˉ 2 ) , ( x ˉ 1 ∧ x 2 , u ˉ 1 ∧ u 2 ) , ( x 1 , u 1 ∧ u 2 )
is a UDL that computes precisely the strategy s s s .
At the end we want to construct a QCDCL refutation of Φ \Phi Φ consisting of at most 3 trails that are not CDCL steps, and for this we follow the proof of Theorem 49.12 In stage 1 we assign X X X according to γ 1 \gamma_1 γ 1 and then U U U according to μ 1 \mu_1 μ 1 . By the construction in the proof, the first trail T 1 , 1 \mathcal{T}_{1,1} T 1 , 1 looks as follows – decisions are blue, propagated literals are white with their antecedent clauses below, and the conflict is red:
Figure 1 (PDF p. 36).
T 1 , 1 = ( x ˉ 1 ; x ˉ 2 ; u ˉ 1 , y 1 , y ˉ 2 , ⊥ ) \mathcal{T}_{1,1} = (\bar{x}_1;\bar{x}_2;\bar{u}_1,y_1,\bar{y}_2,\bot) T 1 , 1 = ( x ˉ 1 ; x ˉ 2 ; u ˉ 1 , y 1 , y ˉ 2 , ⊥ )
Here the conflict occurs even before u 2 u_2 u 2 is decided, since the clause C 2 C_2 C 2 becomes the unit clause ( u 2 ) (u_2) ( u 2 ) , which is reduced to the empty clause. We do not need to perform any CDCL steps and can immediately learn the clause E 1 = x 1 ∨ x 2 = ¬ D 1 E_1 = x_1 \lor x_2 = \neg D_1 E 1 = x 1 ∨ x 2 = ¬ D 1 by resolving C 2 C_2 C 2 with the antecedents C 3 C_3 C 3 and C 1 C_1 C 1 of y ˉ 2 \bar{y}_2 y ˉ 2 and y 1 y_1 y 1 and reducing the universal literals u 1 u_1 u 1 , u 2 u_2 u 2 . In general, we may need some CDCL steps before we can learn a subclause of ¬ D 1 \neg D_1 ¬ D 1 , but these CDCL steps would be ignored for the proof size anyway. We continue with stage 2 and the trail T 2 , 1 \mathcal{T}_{2,1} T 2 , 1 .
Figure 4 (PDF p. 37).
T 2 , 1 = ( x ˉ 1 , x 2 ; u ˉ 1 , y 1 , y ˉ 2 , ⊥ ) \mathcal{T}_{2,1} = (\bar{x}_1, x_2; \bar{u}_1, y_1, \bar{y}_2, \bot) T 2 , 1 = ( x ˉ 1 , x 2 ; u ˉ 1 , y 1 , y ˉ 2 , ⊥ )
where x 2 x_2 x 2 is propagated by the learned clause E 1 E_1 E 1 , which is consistent with γ 2 \gamma_2 γ 2 . This time the conflict occurs with C 6 C_6 C 6 . Resolving C 6 C_6 C 6 with the antecedents C 3 C_3 C 3 and C 1 C_1 C 1 of y ˉ 2 \bar{y}_2 y ˉ 2 and y 1 y_1 y 1 and applying reduction, we get E 2 = x 1 ∨ x ˉ 2 = ¬ D 2 E_2 = x_1 \lor\bar{x}_2 = \neg D_2 E 2 = x 1 ∨ x ˉ 2 = ¬ D 2 , again without any CDCL steps. Finally, stage 3 uses the trail T 3 , 1 \mathcal{T}_{3,1} T 3 , 1 :
Figure 5 (PDF p. 37).
T 3 , 1 = ( x 1 ; x ˉ 2 ; u 1 , y 1 , y ˉ 2 , ⊥ ) \mathcal{T}_{3,1} = (x_1; \bar{x}_2; u_1, y_1, \bar{y}_2, \bot) T 3 , 1 = ( x 1 ; x ˉ 2 ; u 1 , y 1 , y ˉ 2 , ⊥ )
where the conflict with C 5 C_5 C 5 occurs before u 2 u_2 u 2 is decided. We resolve C 5 C_5 C 5 with the antecedents C 3 C_3 C 3 and C 4 C_4 C 4 of y ˉ 2 \bar{y}_2 y ˉ 2 and y 1 y_1 y 1 , apply reduction and get E 3 = ( x ˉ 1 ) = ¬ D 3 E_3 = (\bar{x}_1) = \neg D_3 E 3 = ( x ˉ 1 ) = ¬ D 3 . In general, it is not guaranteed that we learn the full clauses ¬ D i \neg D_i ¬ D i . Instead, we can only guarantee to learn a subclause of the current ¬ D i \neg D_i ¬ D i (or to find such a clause among the clauses already present). Now the clauses E 1 , E 2 , E 3 E_1, E_2, E_3 E 1 , E 2 , E 3 over X = { x 1 , x 2 } X = \{x_1, x_2\} X = { x 1 , x 2 } form an unsatisfiable CNF, because ⋁ i = 1 3 ¬ D i \bigvee_{i=1}^{3} \neg D_i ⋁ i = 1 3 ¬ D i is unsatisfiable. In the final trail V 1 = ( x ˉ 1 , x 2 , ⊥ ) \mathcal{V}_1 = (\bar{x}_1, x_2, \bot) V 1 = ( x ˉ 1 , x 2 , ⊥ ) , the literals x ˉ 1 \bar{x}_1 x ˉ 1 and x 2 x_2 x 2 are propagated by E 3 E_3 E 3 and E 1 E_1 E 1 , and the conflict occurs with E 2 E_2 E 2 . This trail is a CDCL step, and we learn the empty clause ( ⊥ ) (\bot) ( ⊥ ) by resolving E 2 E_2 E 2 with E 1 E_1 E 1 and E 3 E_3 E 3 . In general, we might need several CDCL steps at this point.
For the QCDCL refutation ι = ( T 1 , 1 , T 2 , 1 , T 3 , 1 , V 1 ) \iota= (\mathcal{T}_{1,1}, \mathcal{T}_{2,1}, \mathcal{T}_{3,1}, \mathcal{V}_1) ι = ( T 1 , 1 , T 2 , 1 , T 3 , 1 , V 1 ) we have constructed, we conclude ∣ ι ∣ − ∣ ι ∣ C D C L = ∣ T 1 , 1 ∣ + ∣ T 2 , 1 ∣ + ∣ T 3 , 1 ∣ = 15 ≤ 3 ⋅ ∣ var ( Φ ) ∣ |\iota| - |\iota|_{\mathrm{CDCL}} = |\mathcal{T}_{1,1}| + |\mathcal{T}_{2,1}| + |\mathcal{T}_{3,1}| = 15 \le3 \cdot|\operatorname{var}(\Phi)| ∣ ι ∣ − ∣ ι ∣ CDCL = ∣ T 1 , 1 ∣ + ∣ T 2 , 1 ∣ + ∣ T 3 , 1 ∣ = 15 ≤ 3 ⋅ ∣ var ( Φ ) ∣ .
Conclusion
We have proven the possibly surprising result that, in a theoretical setting in which SAT calls and purely propositional CDCL steps are not counted, the basic version of CAQE (without failed assumptions) is weaker than QCDCL on Σ 3 b \Sigma_3^b Σ 3 b formulas (after a simple preprocessing step) and weaker than Q-resolution on all formulas of bounded quantifier alternation (cf. Figure 1), while in practice CAQE outperforms almost all competing QBF solvers. This indicates that optimisations such as failed assumptions might play an important role for the performance of the solver. In future work, our lower and upper bound techniques could be used to prove separations between different variants of CAQE formally, and hence justify the use of modifications on a proof complexity level. Furthermore, the fact that we discarded propositional hardness in both approaches may be a very relevant factor in comparing these two solvers. It can be assumed that direct usage of a SAT solver is a huge advantage for CAQE, while QCDCL as an extension of CDCL utilises SAT solving in a rather intricate and hidden way. Given the relatively simple setup of CAQE (Figure 2 ) this also means that CAQE can at least in principle use the latest state-of-the-art SAT solver without any changes to its implementation, while QCDCL does not have access to the latest SAT technology without major implementation effort.
We have also shown that, discounting propositional hardness, the proof size of CAQE{sup}Core` is essentially determined by the number of optimal moves, counted up to their effect on the clauses. Further research into this direction may have the potential to not only improve the performance of CAQE itself, but also lead to decision heuristics for QCDCL.